Knowledge Evolution in the Mobile Industry via Embedding-Based Topic Growth and Typology Analysis

Abstract

The mobile industry has experienced long-run changes in its knowledge structure, including identifiable transition points observable through embedding-based semantic analysis. Using abstracts from 86,674 mobile industry publications published between 2005 and 2024, we embed documents with SPECTER2, build year-specific embedding distributions, and derive knowledge regimes by combining change-point detection with inter-year distribution distances. We then extract regime-specific topics via clustering and reconstruct topic lineages by aligning topic similarities to classify inheritance, differentiation, convergence, and disappearance. The analysis delineates three regimes spanning 2005 to 2012, 2013 to 2019, and 2020 to 2024, with pronounced transitions around 2012 to 2013 and 2019 to 2020. Regime 1 centers on foundational technologies such as wireless communication, power, sensors, and reliability. Regime 2 expands toward platforms, apps, and data analytics alongside cross-domain convergence. Regime 3 is characterized by strengthened 5G operations and data-driven services, together with the independent rise in policy, governance, and regulation topics. Transitions reflect recombination built on inherited knowledge rather than abrupt replacement, and post-transition topics display distinct growth typologies by network position and growth pattern. By integrating embedding-based changepoint detection with topic lineage reconstruction, we provide a reproducible account of regime transitions and quantitative evidence to inform the timing of corporate R&D, standard and platform strategies, and policy and regulatory design.

1. Introduction

The mobile industry has shifted from handset-centered competition to platform competition in which operating systems, app stores, and developer ecosystems shape the direction of innovation [1,2]. Even as functional convergence has progressed, the smartphone industry has continued to exhibit persistent product differentiation across manufacturers, and a single dominant design has not fully stabilized [3]. Platform diffusion dynamics interact with ecosystem feedback mechanisms such as network effects, implying that innovative success can be determined not only by device performance but also by the interplay between complementary supply and user scale [4]. App stores have also evolved beyond distribution channels into governance devices that reshape market access and innovation incentives through rule-setting and enforcement, becoming a central agenda in debates on digital governance and regulation [5]. As regulatory regimes such as the European Union Digital Markets Act (DMA) increasingly adopt ex ante constraints on gatekeeper platforms, the institutional environment surrounding app store governance is also changing [6]. This raises the possibility that systematic monitoring of knowledge evolution may help regulators identify technological domains that are becoming infrastructural bottlenecks or gateway-like dependencies before market power becomes fully consolidated.

With the introduction of fifth-generation networks, network computing integration paradigms such as mobile edge computing have expanded, reshaping service architectures and operational modes [7]. Relatedly, the rise in discussions on zero-touch management oriented toward autonomous network operations indicates that operational automation and data-driven decision-making have become core competitive factors in communications infrastructure [8]. The mobile industry thus represents a setting in which technological, platform-ecosystem, and institutional shifts overlap. This convergence creates conditions under which the knowledge system is likely to undergo reallocation and concentration during transition periods [5,8].

In such a rapidly evolving environment, tracing change solely through market outcomes or product launch data tends to rely on indicators observable only after transitions have materialized, making early detection difficult. Scholarly knowledge constitutes an upstream layer in which new concepts, methods, and evaluative frames accumulate before they are reflected in market outcomes, making the scientific literature a more timely signal of industrial change [9]. Although prior work analyzing mobile ecosystem knowledge flows through patent citation networks has illuminated interfirm knowledge movement and structural change, such approaches do not capture how topical content is reorganized at the textual level [10]. In a digital economy where data-driven value creation is increasingly salient, issues such as data access, platform dominance, and policy can become intertwined with the competitive structure of the mobile industry, further widening this gap [11].

Platform policy shifts, such as tracking restrictions in iOS, increasingly need to be interpreted as transitions that combine technological and institutional change [12]. Studies on the relationship between targeting efficiency and privacy in mobile advertising further suggest that constraints on data access can affect performance and competition, supporting the view that data governance changes may intersect with industrial innovation pathways [13]. Despite this rich backdrop, most existing mobile research leaves three gaps unaddressed. First, no study has applied distributional, embedding-based methods to detect transition points in mobile industry knowledge (Gap 1) [9,10]. Second, there is no conservative, size-weighted procedure for reconstructing topic lineages covering inheritance, differentiation, convergence, and disappearance across regime boundaries (Gap 2). Third, no multidimensional growth typology has integrated structural network position with temporal growth pattern to distinguish core architectural reconfiguration from peripheral issue-driven surges (Gap 3). Addressing these gaps requires moving beyond keyword frequencies to examine distributional shifts in document-level semantic space; embedding models in the SPECTER family offer exactly this foundation by representing documents as meaning vectors at scale [14,15].

In domains where technology, platforms, and institutions become rapidly entangled, presenting a static list of topics is insufficient to explain transition periods. Understanding such periods requires reconstructing both the relational structure among topics and their evolutionary pathways across time, how topics persist, split, merge, and disappear.

Distinguishing growth typologies after transitions clarifies which topics move toward the core of the knowledge system, which persist in the periphery, and which surge during specific periods. This motivates an integrated approach that constructs year-specific knowledge distributions from embedding-based document representations, identifies transition points from distributional shifts, and links topic structures before and after transitions into lineages. Mobile industry evolution should not be viewed as the cumulative growth of a single technology; rather, it progresses through the reallocation of knowledge exploration and exploitation, the emergence of new problem frames enabled by knowledge recombination, and interactions with shifting platform rules and institutional environments. Building on this perspective, we interpret regime transitions as structural changes in innovation phases and treat topic lineage events such as inheritance, convergence, and differentiation as observable indicators of knowledge recombination.

Academically, the study offers an integrated framework for explaining long-term knowledge change in the mobile industry from a structural perspective, one centered on distributional transitions and the reconfiguration of topic lineages, rather than the growth of specific technologies. Methodologically, it combines embedding-based transition-point detection with topic alignment-based lineage reconstruction to provide a reproducible procedure for identifying transition periods and characterizing their nature. Practically, the findings distinguish topics that move toward the central axis after transitions from those that persist at the periphery, providing evidence to support corporate R&D prioritization, platform and standard strategies, and the exploration of collaboration and investment opportunities. From a policy perspective, the results offer quantitative implications for the timing of regulatory support and the prioritization of policy agendas, grounded in structural shifts, notably the rise in data, governance, and regulation topics following the fifth-generation transition.

Against this backdrop, the present study addresses three interconnected research questions. RQ1: At what temporal boundaries does the mobile industry knowledge system undergo statistically verifiable distributional transitions, and what auxiliary structural indicators, centroid displacement, and dispersion change characterize the nature of each transition? RQ2: How do individual topics persist, merge, split, emerge, or disappear across regime boundaries, and what weighted contribution criteria enable conservative, reproducible classification of these evolutionary events? RQ3: What two-dimensional typology, combining structural centrality in the topic transition network with temporal growth intensity, best differentiates topics whose growth reflects core architectural reconfiguration from those reflecting peripheral issue-driven surges, and what strategic and regulatory prescriptions follow from each type?

The proposed framework is reproducible, scalable to corpora of 86,000-plus documents, and yields results that are robust across 432 hyperparameter combinations with 97.2% boundary stability, as demonstrated in the sensitivity analysis reported in Section 5. A complementary gradient test across nine (τ and δ) combinations confirms that the six primary core bursty lineages and the recombination-dominant interpretation of regime transitions are fully invariant to parameter variation.

The remainder of this paper is organized as follows. Section 2 reviews the theoretical background and prior studies on the mobile industry knowledge structure, platform ecosystems, transition-point analysis, and topic evolution and lineage reconstruction. Section 3 describes the data collection and preprocessing, document embedding construction, transition-point detection procedure, regime-specific topic extraction, and the methods for topic alignment and growth typology classification. Section 4 presents the results on transition points and regime identification, regime-specific topic structures, and analyses of topic lineages and growth typologies. Section 5 discusses the findings and derives implications from academic, industrial, and policy perspectives. Section 6 concludes by summarizing contributions, discussing limitations, and suggesting directions for future research.

2. Literature Review

2.1. Research on the Mobile Industry

The mobile industry exhibits a multi-sided platform structure that simultaneously matches users and complementors, and the theoretical foundation is two-sided market theory, which emphasizes that cross-side network effects shape market competition and the pace of innovation [16,17,18]. Platforms coordinate participation on both sides through pricing, access, and rules, and they design both value creation and value capture, providing a core analytical lens for understanding the mobile OS app store payment and advertising and developer ecosystem [18]. In such markets, competition can be reconfigured less by share levels per se than by the process through which platform boundaries expand and overlap, and platform envelopment strategies that absorb adjacent markets can rapidly reorder industrial boundaries and innovation trajectories [19].

Platform research has extended into ecosystem theory, strengthening the view that innovation outcomes depend not only on focal firm capabilities but also on interdependence among components and coordination failures [20,21]. Ecosystems are analyzable as structures rather than as simple networks, and accumulated discussions show that competitive advantage varies with which components perform which roles and where bottlenecks arise [21,22]. Industrial platforms can induce external innovation and accelerate ecosystem-level innovation through modularity, interface design, and governance, which have refined platform leadership and orchestration as central strategic themes [23,24].

As platforms grow, their relationships with complementors involve not only collaboration but also competition, and tensions and shifts in innovation incentives can arise when platforms enter complementor domains [25]. Platforms must also choose strategic tradeoffs between openness and control, and governance, rules, and fee structures can structurally reallocate participant behavior within ecosystems [26,27]. Within this setting, antitrust frameworks and competition policy debates for multi-sided platforms require regulatory logics distinct from those of traditional industries, motivating calls for competition policy frames tailored to the digital era [28,29]. Work that systematically integrates the platform competition literature synthesizes these core issues, such as competition, dominance, governance, ecosystems, and complementor strategies, and provides coordinates for research design [30].

Finally, the mobile industry has faced an increased likelihood of transition periods in which technological, platform, and institutional elements change simultaneously, as the generational shift from 4G to 5G has redefined networks as foundational infrastructures for services and operations [31,32]. Edge computing in particular alters constraints related to latency, bandwidth, and service design through tighter coupling of networks and computing, and it is summarized as a driver that can trigger structural change in mobile service architectures and operating modes [33,34,35]. On the operational side, standardization and automation discussions, such as zero-touch service management, indicate that data-driven operational paradigms have emerged as core competitive factors in communications infrastructure [36].

2.2. Bibliometric Studies and Knowledge Flow Research in the Mobile Industry

In domains such as the mobile industry, where technological and service generations turn over rapidly, bibliometrics and science mapping have been widely used to capture the accumulation and diffusion of knowledge quantitatively. From a network perspective, scholarly knowledge can be represented through citation relations among papers and patents, and general theories for analyzing such multiplex networks provide a basis for quantitatively summarizing science and technology knowledge flows [37]. Network analytic methodologies have also been refined to explain how knowledge is organized and connected using structural indicators such as centrality, betweenness, and community structure [38].

Two representative linkage rules in bibliometric networks are bibliographic coupling and co-citation. Bibliographic coupling assumes that two documents are more similar when they share the same references, and there is classical work proposing linkage rules for coupling across documents [39]. Co-citation defines similarity by how frequently two documents are cited together by third documents, and it has become a core tool in the analytic tradition that distinguishes research fronts from intellectual bases to characterize knowledge structures [40]. These linkage rules provide static similarity but reconstructing how knowledge flows unfold along temporal paths requires explicit path extraction.

Main path analysis was proposed to identify the central pathways along which knowledge flows in citation networks, and this approach has continued to develop through improved connectivity-based reconstruction methods [41]. Efficient algorithms and implementations that enable main path analysis on large-scale networks have subsequently diffused across diverse science and technology fields [42]. Studies that reconstruct technological trajectories using patent citation networks have highlighted both the usefulness of main path analysis and limitations, such as missing important nodes and producing overly complex paths, thereby proposing directions for improvement [43]. Related approaches that interpret patent citations in staged ways to quantify inventive progress can also be understood as efforts to jointly explain paths and stages of technological evolution [44]. More recently, studies have proposed improvements to patent-based main path extraction or generalized procedures, further expanding the empirical toolbox for analyzing technological evolution [45]. In the 5G domain as well, research combining patent citation networks with main path analysis has emerged to trace technological development streams, accumulating analytic grounds that connect generational transitions with knowledge flows in the mobile industry [46].

Knowledge structure change can appear not only as paths but also as abrupt surges, and a representative quantitative device for capturing such surges is burst detection. Burst detection algorithms that model the short-term intensification of specific terms or topics in document streams have been used as a methodological basis for detecting trend transitions across many fields [47]. Visual analytic tools that map temporal variation in research fronts and intellectual bases and detect clues of transitions through measures such as centrality-based pivotal points have also become representative approaches supporting the interpretation of knowledge structure change [48].

From an empirical standpoint, these methodologies have been applied to quantify research fronts and hot topics in specific themes such as 5G security and 5G applications [49]. However, in domains like the mobile industry, where industrial, standard, patent, and scholarly knowledge accumulate simultaneously, it is difficult to capture the direction and speed of innovation using a single data source, increasing the need to jointly design patent citation-based network analysis with bibliometric and text analytic approaches. Research monitoring brokerage roles in patent citation networks from an open innovation perspective quantitatively shows that knowledge movement across firms and technologies can be linked to innovation strategy [50]. The conceptual framework of open innovation further emphasizes that innovation activities have shifted from closed internal R&D toward actively combining external knowledge and pathways, and it has been used as a background theory for industry ecosystem analysis [51].

Finally, recent bibliometric analysis has improved substantially in reproducibility and scalability alongside advances in tool ecosystems. VOSviewer has been widely used as a representative tool for visualizing and mapping large-scale bibliometric networks [52]. Theoretical formalizations that seek to integrate mapping and clustering under a unified principle have also been proposed, providing a basis for improving the consistency of tool-based results [53]. Bibliometrix provides an R-based open-source workflow for conducting science mapping, offering an integrated procedure that spans data processing, analysis, and visualization [54]. In sum, quantitative prior research on mobile industry knowledge flows has expanded toward deriving knowledge structures based on bibliographic coupling and co-citation, reconstructing trajectories via citation and patent networks, detecting transition clues through bursts and visual analytics, and building reproducible workflows through tool-based pipelines.

2.3. Methodological Review

To argue for regime shifts in knowledge structures within long-run corpora, it is necessary to detect changes in distributions themselves rather than changes in publication volume or keyword frequency. In multivariate settings, nonparametric methods that estimate multiple change points can directly infer when regimes change from data, providing a core foundation for transition analysis [55]. Change-point detection has also advanced in computational efficiency, and approaches that detect optimal change points with linear cost have been proposed, expanding feasibility for large-scale time-series applications [56]. Reviews synthesizing change-point detection methods emphasize that method choice should depend on data characteristics and objectives, such as offline versus online settings and univariate versus multivariate cases [57].

Quantifying the strength of transitions requires statistics that measure distributional differences, and distance-based statistics such as energy statistics provide a framework that expresses distributional differences as distances [58]. Results showing theoretical connections between distance-based tests and reproducing kernel Hilbert space-based tests provide justification for method selection in distribution testing [59]. Maximum mean discrepancy, formulated as a kernel-based two-sample test, can capture distributional differences broadly and can therefore be used to support the magnitude and direction of candidate transition periods [60].

Topic modeling has become a representative technique for extracting themes from the science and technology literature [61]. Dynamic topic models that explicitly incorporate time have also been proposed, and efforts to model long-term trends have continued [62,63]. More recently, advances in contextual embeddings have reshaped topic construction methods. Pretrained BERT family models substantially improved language representations [64], and pretrained models specialized for scientific text have also been proposed, strengthening the foundation for analyzing science and technology documents [65].

Embedding-based topic construction relies on dimensionality reduction and clustering as core procedures, and UMAP has been widely used to transform high-dimensional semantic space into representations that are suitable for visualization and clustering [66]. HDBSCAN is well-suited to text corpora because it can detect topics of varying sizes and densities without pre-specifying the number of clusters, and it provides practical advantages by separating noise while forming stable clusters [67,68]. BERTopic combines embedding, dimensionality reduction, clustering, and class-based TF-IDF to construct interpretable topics in large-scale text, and it has become a representative implementation of embedding-based topic modeling [69].

Explaining topic evolution requires the perspective that topics change not only by increasing or decreasing but also through events such as emergence, disappearance, differentiation, and convergence. Approaches have been proposed to visually represent topic flows to make complex changes interpretable [70], and work that structurally connects transitions among hierarchical topics provides a logical basis for topic lineage reconstruction [71]. Reviews of the development of topic modeling organize model families and evaluation issues and emphasize the need for procedure design aligned with research purposes such as exploration, tracking, and explanation [72].

2.4. Limitations of Prior Studies and Research Gaps

The review of prior work reveals three methodological gaps that the present study is designed to address. First (Gap 1), existing knowledge flow studies in the mobile industry have relied primarily on patent citation networks or keyword frequency analysis. Methods that directly detect distributional shifts in document-level semantic space remain underdeveloped, limiting the ability to identify transition periods objectively and to capture early signals of structural change [9,10]. Second (Gap 2), topic modeling studies typically present static snapshots of topic structures within a given period, and systematic procedures for reconstructing topic lineages across regime boundaries, integrating inheritance, differentiation, convergence, and disappearance, are lacking. In particular, conservative classification approaches that use weighted contributions to avoid over-identifying merge and split events have not yet been established. Third (Gap 3), topic growth is most commonly assessed using single indicators, such as document count increases, and multidimensional typology frameworks that jointly consider structural position in transition networks, but temporal growth patterns remain absent. This restricts the ability to distinguish reconfiguration surges at the knowledge core from issue-driven surges at the periphery.

These three gaps collectively point to a deeper epistemological limitation in prior mobile industry knowledge studies: the inability to distinguish between gradual cumulative growth within a stable paradigm and punctuated structural reconfiguration that reorients the knowledge system toward a new problem center. This distinction has direct implications for technology management: firms that cannot differentiate between incremental and architectural knowledge change are prone to underinvest during bursty reconfiguration phases and overinvest in mature peripheral domains. The present study’s integrated framework, combining distributional regime detection, conservative lineage reconstruction, and multidimensional growth typology, is designed to provide precisely this discriminatory capacity at the level of a large-scale longitudinal corpus.

A fourth gap is also identified. No prior bibliometric study has explicitly connected embedding-based topic centrality to the cospecialized assets framework [73], linked topic spike patterns to real options theory [74], or mapped peripheral bursty trajectories onto the disruptive innovation model [75]. Bridging this theoretical gap is a secondary contribution of the present study, pursued through the three formal propositions derived in Section 5.1 and the typology-specific strategic prescriptions in Section 5.2. Together, these contributions position the study at the intersection of computational bibliometrics and strategic management of technology, advancing both the methodological toolkit and the theoretical vocabulary available for analyzing knowledge evolution in high-velocity industries.

To address these gaps, this study proposes a methodological framework that integrates E-Divisive and MMD-based regime detection, topic alignment that combines similarity with weighted contribution, and a two-by-two typology that jointly considers structural position and growth patterns.

3. Materials and Methods

The purpose of this study is to empirically examine how scholarly knowledge in the mobile industry accumulates over time, when it is structurally reconfigured, and which thematic strands subsequently converge toward the core or persist at the periphery. To this end, we construct a large-scale longitudinal text corpus based on Web of Science abstracts published between 2005 and 2024 and implement an integrated analytical pipeline consisting of four steps, as summarized in Figure 1. First, we quantify year-specific knowledge distributions by generating document embeddings. Second, we detect transition periods and segment the observation window into regimes based on distributional shifts. Third, we construct regime-specific topic structures. Fourth, we align topics across regimes to derive topic lineages and classify topic growth typologies by jointly considering structural position and temporal growth patterns.

3.1. Data Collection and Preprocessing

3.1.1. Data Collection

This step constructs the literature corpus required for the analysis. Building on prior studies, we first organized the mobile industry literature into analytical categories and then developed a structured block-based search query accordingly. As shown in

Table 1. Web of Science search formula.

Category	Formula
Devices, OS, industry, and ecosystem	TS_BLOCK1 = ((smartphone OR ‘smart phone’ OR ‘mobile phone’ OR cellphone OR ‘cell phone’ OR ‘cellular phone’ OR handset OR ‘mobile device’ OR ‘feature phone’ OR tablet OR ‘mobile terminal’ OR iPhone OR Android OR iOS OR Symbian OR ‘Windows Phone’ OR BlackBerry) AND (‘mobile industr’ OR ‘smartphone industr’ OR ‘handset industr’ OR ecosystem OR ‘value chain’ OR ‘supply chain’ OR manufacturing OR production OR platform OR ‘app store’ OR ‘business model’ OR market OR vendor OR OEM OR brand OR ‘market share’ OR competition OR strateg OR pricing OR ‘intellectual property’ OR patent OR standard OR ‘3GPP’ OR ‘LTE’ OR ‘5G’))
Telecom operators, performance metrics, and policy	TS_BLOCK2 = (((mobile OR cellular OR wireless) NEAR/3 (operator OR carrier OR telecom OR ‘service provider’ OR MNO OR MVNO)) AND (ARPU OR churn OR tariff OR spectrum OR licensing OR regulation OR ‘network sharing’ OR roaming OR ‘base station’ OR RAN OR ‘core network’ OR ‘VoLTE’ OR ‘5G NR’ OR ‘non-standalone’ OR ‘standalone’ OR vendor))
Components, semiconductors, display, supply chain, and production	TS_BLOCK3 = (((smartphone OR handset OR ‘mobile device’) NEAR/3 (chip OR SoC OR baseband OR modem OR ‘RF front end’ OR display OR OLED OR AMOLED OR LTPO OR ‘camera module’ OR ‘image sensor’ OR CIS OR battery OR charger OR PMIC OR memory OR DRAM OR NAND OR ‘touch panel’ OR ‘cover glass’ OR ‘Gorilla Glass’ OR PCB OR packaging OR SiP OR foundry OR fab OR TSMC OR ‘Samsung Foundry’ OR Qualcomm OR MediaTek OR Sony)) AND (market OR vendor OR ‘supply chain’ OR manufacturing OR production OR capacity OR ‘lead time’ OR shortage))
Mobile apps, services, and ecosystem	TS_BLOCK4 = ((mobile NEAR/3 (app OR ‘app store’ OR ‘mobile service’ OR ‘mobile payment’ OR ‘m-payment’ OR ‘mobile banking’ OR fintech OR ‘ride-hailing’ OR ‘social media’ OR ‘messaging app’)) AND (market OR monetization OR platform OR ecosystem OR competition OR pricing OR adoption OR diffusion))
Standards, SEP, patents, licensing, and royalties	TS_BLOCK5 = ((‘3GPP’ OR ‘Release 15’ OR ‘Release 16’ OR ‘Release 17’ OR ‘LTE’ OR ‘LTE-Advanced’ OR ‘5G’ OR ‘5G NR’ OR ‘New Radio’ OR ‘6G’ OR ETSI OR ‘IMT-2020’ OR ‘IMT-2030’ OR ‘standard-essential patent’ OR SEP OR FRAND) AND (industry OR market OR standard OR patent OR litigation OR licensing OR royalty))

, the target set is restricted to scholarly articles related to the mobile industry and the smartphone industry, identified through this query. To ensure reliability and reproducibility, we use the Web of Science Core Collection as the data source. The analysis period spans 2005 to 2024 to capture long-term dynamics in which technological generational shifts and industrial restructuring accumulate over time. We restrict the document type to article and extract the bibliographic information required for subsequent analyses, including authors, titles, abstracts, keywords, journals, affiliations, and citation metadata. This process yields a final dataset of 86,674 records.

3.1.2. Preprocessing

The preprocessing stage refines abstract text into an analyzable format. We first organized documents by publication year to construct a longitudinal corpus and then removed function words with low semantic contribution, such as articles, prepositions, and conjunctions, as well as generic terms that reduce discriminative power.

To mitigate concept fragmentation caused by spelling and wording variations, we applied standardization for key terms and handled synonyms by unifying equivalent expressions, such as platform and ecosystem, and standard and standardization. These steps minimize noise in subsequent embedding and clustering procedures and improve the interpretability of the resulting topics.

3.2. Regime Identification and Segmentation

3.2.1. Document Embedding

This section converts longitudinally accumulated paper abstracts into embedding-based numerical vectors so that semantic distributions can be compared across years. Because raw text is not directly suitable for inter-year comparison or for measuring distributional shifts, each document is mapped to a vector representation that preserves semantics, enabling similarity and distance-based analyses in a shared coordinate space.

The time-ordered abstract corpus is encoded using the SPECTER2 model. SPECTER2 is a pretrained model that learns scholarly document representations using citation contexts and is well-suited to capturing semantic proximity in science and technology texts [14,15]. Each abstract is transformed into a fixed-length vector of 768 dimensions, such that semantically similar documents are located closer to one another in the embedding space. The choice of SPECTER2 over alternative embedding models rests on two considerations. First, SciBERT [64], while trained on scientific text, is designed and optimized primarily for token-level and sentence-level discriminative tasks such as named entity recognition and relation extraction [64]; its pretraining objective (masked language modeling on scientific corpora) does not incorporate any document-level relatedness signal, making it less suited to the distributional similarity task central to this study. As the original SPECTER paper explicitly notes, models such as BERT and SciBERT are targeted towards token- and sentence-level training objectives and do not leverage information on inter-document relatedness, which limits their document-level representation power [14]. The same limitation applies to general-purpose language models such as BERT [63]. Competing scientific document embedding models such as SciNCL [76] address the document-level representation problem through citation-graph-based contrastive learning and perform comparably to SPECTER on the SciDocs benchmark but do not offer the multi-task adapter framework of SPECTER2 that enables task-format-specific embeddings. Second, on the SciRepEval benchmark, the first comprehensive multi-format evaluation covering 24 tasks across classification, regression, ranking, and search, SPECTER2 Adapters outperform both SPECTER and SciNCL by more than two absolute points on average across out-of-training tasks [15]. This external benchmark validation confirms that SPECTER2 provides superior document-level representations compared to available alternatives for the proximity-based distributional analysis employed in this study. To assess whether the substantive conclusions are sensitive to embedding model choice, a robustness check was conducted by re-encoding a stratified 20% subsample of the corpus (17,335 abstracts, proportionally drawn across years) using SciBERT and repeating the regime detection and lineage reconstruction pipeline. The E-Divisive procedure identified the same two change points (2012–2013 and 2019–2020) in 100% of permutation runs on the SciBERT-encoded subsample, and the classification of the six primary core bursty lineages was identical under both encoders. These results confirm that the main findings are not an artifact of the specific embedding model chosen.

To reduce the influence of vector magnitude and to improve the stability of distributional comparisons across years, L2 normalization is applied. This enables consistent estimation of distribution-based statistics required for transition analysis, such as shifts in the centroid and changes in dispersion. All analyses were implemented in Python 3.10 using Google Colab.

3.2.2. Regime Shift Detection

This section quantitatively identifies regime shifts in which the mobile industry knowledge structure changes beyond gradual variation and exhibits a distinct distributional character. In this study, a regime shift is defined as a point at which continuity between consecutive year-level embedding distributions weakens substantially and the knowledge system is reorganized into a different structure. Specifically, for a given year t, a regime shift is assumed when the embedding distribution before t and the embedding distribution after t differ in a statistically meaningful way.

Regime shift detection is conducted in two stages. First, to generate candidate boundaries, a multiple change-point detection procedure is applied to the embedding time series. Document embeddings are arranged by year, and an E-Divisive-based nonparametric method is used to detect boundary indices at which the underlying distributions change significantly. The number of regimes is therefore not predetermined. Instead, it is determined endogenously by the number of statistically significant change points detected in the annual embedding sequence. Statistical significance is evaluated using a permutation-based test with a predefined significance threshold of p < 0.05, and only boundaries satisfying this criterion are retained as candidate regime shifts. This stage produces a candidate set of years that may mark distributional reconfiguration and serves as an exploratory step for identifying strong transitions in the embedding time series.

Second, to confirm the substantive importance of the candidate boundaries, a time series of distributional distances based on maximum mean discrepancy is constructed. For each adjacent year pair, t and t + 1, the distributional difference is computed as the MMD between the embedding sets of the two years, forming a year-pair change series. Boundaries are then prioritized around peak segments where the change magnitude increases sharply. Years for which E-Divisive candidates and MMD peaks are jointly observed are treated as high-likelihood transition points. To avoid excessive sensitivity to a single detection procedure or local fluctuation, regime shifts are finalized by cross-validating the stage-one candidates against the stage-two peak structure.

Finally, two auxiliary indicators are used to assess the structural plausibility of the detected shifts. Centroid shift is computed to measure how far the mean location of the distribution moves across a boundary and to evaluate whether an increase in MMD is driven by translation of the overall topical center. Dispersion change is also computed to examine whether the shift is associated with expansion or contraction of distributional spread, thereby supporting interpretation in terms of knowledge diversification or convergence. These indicators extend the analysis beyond statistical detection by enabling characterization of whether the detected boundaries reflect genuine structural reconfiguration rather than temporary sampling fluctuation. All analyses are implemented in Python 3.10 using Google Colab.

3.2.3. Regime Segmentation

This section segments the full observation period into multiple regimes using the finalized transition years as boundaries, thereby establishing comparable period-specific structures for subsequent analysis. Using the detected boundary years, the 2005 to 2024 corpus is partitioned into contiguous intervals, and each interval is treated as a relatively homogeneous knowledge system. Each regime then becomes the unit of analysis for constructing topic structures and analyzing topic dynamics. After segmentation, regime-specific document sets, embedding distributions, and topic clusters can be constructed independently, and structural differences and inheritance relationships before and after transitions can be systematically compared in later steps. All analyses were implemented in Python 3.10 using Google Colab.

3.3. Topic Structure Construction

3.3.1. Topic Clustering per Regime

This section derives topics and constructs regime-specific topic structures by identifying density patterns in document embeddings within each segmented regime. Regime-level topic clustering reveals how studies produced within a given period are organized into subtopic sets and provides the basic units for cross-regime topic alignment and topic dynamics analysis.

Document embeddings obtained in Section 3.2 are split using the regime boundaries to form an embedding set for each regime. HDBSCAN, a density-based clustering method, is then applied to cluster each regime embedding distribution into topics. Because HDBSCAN forms clusters based on density variation, the number of topics is not fixed in advance and can be determined flexibly by the data structure within each regime. HDBSCAN also separates noise points, allowing documents that do not stably belong to any topic to be treated as outliers, which improves topic homogeneity.

To enhance the interpretability of the derived topic structures, regime-specific topics are projected into a two-dimensional space using UMAP for visualization. Because UMAP aims to preserve neighborhood relations from the high-dimensional embedding space, it is used to inspect relative distances among topics, the degree of topic separation, and regime-level structural differences in an intuitive way.

Finally, representative keywords are computed for each topic by aggregating the abstract texts of documents assigned to the topic. Class-based TF-IDF is applied to extract highly weighted terms that characterize each topic, providing a basis for topic labeling and semantic interpretation.

3.3.2. SPECTER-Based Topic Representation

This section represents regime-specific topics derived by HDBSCAN in a consistent manner in the SPECTER embedding space so that topic meanings can be summarized quantitatively and used in downstream steps, including cross-regime topic alignment and topic dynamics analysis. Topic-level representations are constructed by summarizing multiple document embeddings into a single representative vector while also producing interpretable descriptors such as representative documents and keywords.

For each topic, a representative vector is defined to capture the central tendency of document embeddings within the topic. Specifically, the representative vector of topic k is computed as the centroid, the mean of the document embeddings assigned to the topic. This represents the topic as a single point in the SPECTER semantic space and serves as a key input for topic similarity computation, cross-regime matching, and lineage reconstruction. Because centroid-based representations reflect shared meaning within a topic, they provide comparability even when the number and distribution of topics differ across regimes.

Because a representative vector alone is not directly interpretable, representative documents and representative keywords are additionally identified for each topic. The representative document is defined as the document whose embedding is closest to the topic centroid, serving as an example that best captures the topic content. Representative keywords are obtained by aggregating abstracts within each topic and extracting the top terms using class-based TF-IDF weighting. Each topic is thus described by a three-component representation consisting of the centroid vector, a representative document, and a representative keyword set, enabling both quantitative comparison and qualitative interpretation.

To present topic structures intuitively, a topic map is constructed by projecting topic centroid vectors into two-dimensional space using UMAP. This visualization illustrates relative distances and cluster structures among topics and supports inspection of regime-specific topic configurations and topic positions, such as central versus peripheral locations. However, UMAP is used only for visualization, while quantitative comparison and alignment are performed in the original high-dimensional embedding space based on similarities among topic centroids.

In summary, the SPECTER-based topic representation provides a standardized topic-level expression by defining topics quantitatively through centroid vectors, ensuring interpretability through representative documents and class-based TF-IDF keywords, and visualizing structures through a UMAP-based topic map. This representation supports subsequent cross-regime topic alignment and growth typology analysis.

3.4. Topic Dynamics and Growth Typology

3.4.1. Cross-Regime Topic Alignment

This section temporally connects topics that are independently derived within each regime, reconstructs topic inheritance relationships, and systematically identifies evolutionary events such as birth, death, merge, split, and recombination. Because regime-specific topics are constructed from period-specific data distributions, they do not share a common topic index system. An explicit alignment procedure is therefore required to generate linkage edges by matching topic representative vectors across adjacent regimes.

Semantic continuity between topic k in regime T and topic l in regime T plus 1 is assessed using similarity between topic representative vectors. Each topic is summarized by two attributes: a representative vector defined as the centroid of document embeddings within the topic and a topic size defined as the number of documents assigned to the topic. For all topic pairs k and l across adjacent regimes, a cosine similarity matrix is computed to form candidate links.

To prevent excessive link creation, three criteria are applied sequentially to confirm one-to-one inheritance relationships. First, a similarity threshold τ is introduced so that only pairs satisfying sim(k, l) ≥ τ are retained as candidates. To determine τ in a data-driven manner, a permutation-based null distribution is constructed, and τ is set automatically to exceed similarity levels expected under random matching. In the present study, this yields τ = 0.7. Second, a margin criterion is used to ensure that the best match is sufficiently dominant. Specifically, for each topic k, the margin is computed as the difference between the highest and the second-highest similarity scores, margin(k) = best(k) − second(k), and one-to-one inheritance is confirmed only when margin(k) ≥ δ. The value of δ is set empirically based on the margin distribution. In the present study, this yields δ = 0.03. Both parameter choices remain robust under alternative values. Third, to reduce misalignment caused by one-sided matching, a mutual top N condition is applied. A link is accepted only when the best candidate l for topic k also includes k among its top N candidates. As a result, a one-to-one inheritance link is defined as a pair that jointly satisfies sim ≥ τ, margin ≥ δ, and the mutual top N condition.

To verify that the substantive findings are not artifacts of the specific parameter values chosen, a gradient sensitivity test was conducted by crossing three levels of τ (0.6, 0.7, 0.8) with three levels of δ (0.01, 0.03, 0.05), yielding nine parameter combinations in total. For each combination, topic alignment was re-executed across both regime transitions (R1→R2 and R2→R3), and three outcomes were tracked: (i) the total number of confirmed one-to-one inheritance links, (ii) the number of reconstructed cross-regime lineages of length ≥ 2, and (iii) the proportion of the 30 focal lineages retaining identical evolutionary event labels relative to the baseline specification (τ = 0.7, δ = 0.03). The results are summarized in

Table 2. Event-classification stability of 30 focal lineages under nine (τ and δ) combinations.

τ\δ		Margin Threshold (δ)
		δ = 0.01	δ = 0.03	δ = 0.05
Similarity Threshold (τ)	τ = 0.6	29/30 (96.7%)	28/30 (93.3%)	27/30 (90.0%)
	τ = 0.7	28/30 (93.3%)	30/30 (100%)	28/30 (93.3%)
	τ = 0.8	27/30 (90.0%)	26/30 (86.7%)	25/30 (83.3%)

.

The six core bursty lineages central to the main findings (L6, L10, L17, L20, L28, and L30) are reconstructed identically across all nine combinations, as their underlying cosine similarity values range from 0.81 to 0.94, well above τ = 0.8. The two anchor core persistent lineages (L2, L7) are similarly invariant (similarities ≥ 0.85). Sensitivity is confined to peripheral lineages whose inheritance similarities fall near the τ boundary (0.68–0.73): under τ = 0.8, three such lineages are reclassified as births, while under τ = 0.6, two marginal links are admitted, generating one additional split event. These boundary-level changes do not alter regime-level characterizations or the recombination-dominant interpretation of transitions. The margin threshold δ exerts a smaller effect (link count variation ≤ 4% within any fixed τ level), confirming that the dominance condition is not the binding constraint. To further verify robustness at the regime segmentation level, the proportion of topics classified as born (In = 0) at each transition was tracked across all nine combinations as a proxy for apparent discontinuity. At the R1→R2 transition, the birth proportion ranges from 18.2% (τ = 0.6) to 27.3% (τ = 0.8), compared to 22.7% at baseline. At the R2→R3 transition, the range is 20.0% to 30.0%, compared to 25.0% at baseline. In all cases, inheritance remains the dominant event type across both transitions, and the qualitative characterization of each transition as recombination-dominant rather than replacement-driven is preserved regardless of parameter choice.

After confirming one-to-one inheritance links, the remaining connections are further interpreted to classify evolutionary events. Topic birth is defined for a topic l in regime T plus 1 when no valid incoming link from the previous regime exceeds the threshold, which corresponds to In(l) = 0. Conversely, topic death is defined for a topic k in regime T when no valid outgoing link to the next regime exceeds the threshold, which corresponds to Out(k) = 0. However, classifying many-to-one or one-to-many patterns as merge or split solely because the number of links is at least two risks over-identification. To address this, the study introduces weights that reflect not only similarity but also topic size and confirms events based on the extent to which a source topic explains a target topic.

As shown in Equation (1), the weighted contribution from topic k in regime T to topic l in regime T plus 1 is defined as w(k, l) = S(k, l) × n(k), where S denotes cosine similarity and n(k) denotes topic size. The incoming share of topic k to topic l is then computed, as shown in Equation (2), as $shar{e}_{in}$(k → l) = w(k, l)/Σ_{k′ ∈ In(l)} w(k′, l). Because this share is a relative contribution based on similarity multiplied by size rather than similarity alone, it enables a more conservative determination of whether a target topic is genuinely formed through convergence.

$$w(k, l)=S(k, l)\times n\left(k\right)$$

(1)

$$shar{e}_{in}(k\to l)={\displaystyle \frac{w(k, l)}{\mathsf{\Sigma }\_\{k^{\prime }\in In(l\left)\right\} w(k^{\prime }, l)}}$$

(2)

A merge is defined as a case in which a topic l in regime T plus 1 is formed through substantive contributions from multiple topics in regime T. To confirm a merge, two conditions are imposed. First, at least the top two $shar{e}_{in}$ values must each be greater than or equal to β. Second, the cumulative sum of the top contributions must be greater than or equal to γ. For example, β can be set to 0.20 and γ to 0.70. In other words, a topic is classified as a merge only when at least two prior topics each contribute at a meaningful level and jointly account for most of the target topic.

A split is defined as the inverse of a merge. Specifically, a split occurs when a topic k in regime T branches into multiple topics in regime T plus 1 through substantive outgoing contributions. To quantify branching, the outgoing share is computed, as shown in Equation (3), as $shar{e}_{out}$(k → l) = w(k, l)/Σ_{l′ ∈ Out(k)} w(k, l′). The confirmation criteria for split are set symmetrically to those for merge. A topic k is classified as a split when at least the top two $shar{e}_{out}$ values are each greater than or equal to β and the cumulative sum of the top contributions is greater than or equal to γ.

$$shar{e}_{out}(k\to l)={\displaystyle \frac{w(k, l)}{\mathsf{\Sigma }\_\{l^{\prime }\in Out(k\left)\right\} w(k, l^{\prime })}}$$

(3)

In this study, knowledge recombination is defined as the structural reorganization of topics across regime boundaries through convergent recombination (merge) or divergent recombination (split) [77,78]. The degree of recombination is operationalized through two existing measures: (1) the number of participating topics N, which captures the breadth of knowledge sources involved in each event, and (2) the weighted share values $shar{e}_{in}$ and $shar{e}_{out}$ defined in Equations (2) and (3), which reflect the relative contribution of each source. Together, N and the share distribution characterize both the scale and balance of each recombination event.

In summary, this section first confirms similarity-based one-to-one inheritance links in a conservative manner. It then introduces share measures that incorporate both similarity and topic size to interpret the remaining multiple links while avoiding over-identification of merge and split events. Finally, it classifies topic evolution using consistent rules that also cover birth and death. This design enables the growth typology analysis in the subsequent section to be conducted not at the level of isolated topics but at the level of temporally connected topic lineages.

3.4.2. Topic Growth Typology

This section presents a method for classifying growth typologies by combining the structural roles and temporal growth patterns of topic lineages constructed from cross-regime topic alignment. Even when lineages exhibit similar growth trajectories, their roles within knowledge flows may differ. Conversely, lineages that occupy central positions may still display distinct growth dynamics. Accordingly, this study adopts a two-by-two typology that integrates network-based structural indicators with time-series-based growth indicators rather than reducing topic growth to a single metric.

The unit of analysis is not an isolated topic within a single regime but a topic lineage defined as a chain of topics connected across adjacent regimes through the alignment procedure described in Section 3.4.1. Each lineage may undergo continuation after birth, experience transformations such as merge, split, or recombination, or terminate through death. Growth typology classification is conducted by comparing where each lineage is positioned in the knowledge flow and how its scale changes over time. Lineages that disappear during the knowledge flow or those that are newly born are excluded from the growth typology analysis.

The typology is defined along two axes. Structural position refers to the centrality of a lineage within the topic transition graph. A lineage-level structural score, struct_score, is computed based on centrality measures such as PageRank, degree, and k core. To remove scale differences and enable relative comparison across lineages, the structural score is converted into a percentile rank and normalized to the 0 to 1 range, which is used as the X-axis value. Higher values indicate a more central, core position with stronger linkage and brokerage roles, whereas lower values indicate a more peripheral position and a higher likelihood of being locally bounded.

Growth pattern captures the temporal expansion dynamics of a lineage. A growth indicator is computed from the lineage-level size time series, such as regime-level document counts or topic size changes. In particular, a spike-based measure is used to distinguish bursty growth from gradual accumulation by capturing whether rapid increases occur at specific points in time. The growth indicator is also converted into a percentile rank and normalized to the 0 to 1 range, which is used as the Y-axis value. Higher values indicate stronger bursty growth characterized by short-term surges, whereas lower values indicate persistent accumulation or stable maintenance without abrupt fluctuations.

Because both the X and Y axes are rank-based measures in the 0 to 1 range, quadrant boundaries are defined using the median threshold of 0.5. Lineages are thus classified into four types: peripheral persistent for X below 0.5 and Y below 0.5, peripheral bursty for X below 0.5 and Y at least 0.5, core persistent for X at least 0.5 and Y below 0.5, and core bursty for X at least 0.5 and Y at least 0.5. Peripheral persistent lineages represent subtopics that accumulate stably within a limited scope. Peripheral bursty lineages are structurally peripheral but exhibit sharp attention spikes at specific times. Core persistent lineages correspond to foundational themes that accumulate over long periods in the center of the knowledge flow. Core bursty lineages represent themes that grow rapidly in the core and emerge as central domains after transition periods.

Finally, the classification results are visualized using an X–Y scatter plot of structural position rank versus growth pattern rank to show where each lineage is located among the four types and to compare distributions across types. By jointly considering temporal change and network roles, this framework provides a systematic basis for explaining the formation, diffusion, and reconfiguration mechanisms of core themes in mobile industry knowledge flows.

4. Results

4.1. Regime Identification and Segmentation Results

4.1.1. Regime Identification Results

Table 3. E-Divisive analysis results.

Segment	Start	End
1	2005	2012
2	2013	2019
3	2020	2024

reports the regime intervals identified through E-Divisive-based change-point detection. Importantly, the number of regimes was not imposed in advance. Rather, the algorithm detected statistically meaningful boundaries in the annual embedding sequence, which partitioned the full observation period into three regimes: Regime 1, spanning 2005 to 2012; Regime 2, spanning 2013 to 2019; and Regime 3, spanning 2020 to 2024. These results indicate that the embedding time series contains multiple discontinuities at which the distribution changes, suggesting that the knowledge structure is reorganized into distinct configurations, particularly around the 2012 to 2013 and 2019 to 2020 transitions.

Figure 2 presents a time series that summarizes embedding distribution differences between adjacent years using maximum mean discrepancy. For each year boundary, the MMD value provides a single measure of how much the overall distribution changes from year t to year t + 1, with larger values indicating greater structural change in the knowledge system. The largest peak is observed for the 2019 to 2020 transition, followed by a comparatively large increase for 2012 to 2013. Among the candidate boundaries suggested by E-Divisive, these results indicate that 2019 to 2020 constitutes the strongest regime shift, while 2012 to 2013 represents the next most salient transition. Because these prominent MMD peaks coincide with the statistically detected E-Divisive boundaries, the resulting three-regime structure is interpreted not as an arbitrary partition but as the outcome of two independently corroborated transition points.

Figure 3 shows year-level publication volumes for mobile industry-related articles from 2005 to 2024, together with the regime intervals determined by the confirmed transition years of 2012 to 2013 and 2019 to 2020. The regimes displayed in the figure—Regime 1, spanning 2005 to 2012; Regime 2, spanning 2013 to 2019; and Regime 3, spanning 2020 to 2024—are defined by boundaries at which discontinuous changes in the document embedding distribution are detected. The publication trend is provided as supplementary evidence to illustrate how the segmentation corresponds to changes in the scale of research production.

In Regime 1, spanning 2005 to 2012, annual publication volume increases gradually from 980 to 2137, indicating a steady accumulation of research output. In Regime 2, spanning 2013 to 2019, publication volume expands from 2564 to 5796, and the growth slope becomes notably steeper, suggesting a transition to an expansion phase in research production. In Regime 3, spanning 2020 to 2024, publication volume jumps sharply to 7921 in 2020 and remains high thereafter, reaching 9170 in 2021, 9692 in 2022, 9804 in 2023, and 10,004 in 2024. Notably, the 2019 to 2020 boundary corresponds to the largest MMD peak and is also associated with a stepwise upward shift in publication volume around the same period.

4.1.2. Regime Validation Results

To further validate the detected regime boundaries, additional indicators are used to assess whether the observed increases in distributional distance reflect substantive structural reconfiguration rather than temporary fluctuation. Figure 4 reports centroid shift, which measures the year-to-year displacement of the mean location of the embedding distribution between consecutive years, normalized for comparability across the observation period. A large centroid shift indicates that the topical center of documents moves collectively in a specific direction, suggesting that directional translation of the distributional center is an important component of the regime transition.

The results show a particularly large centroid shift for the 2019 to 2020 boundary, indicating a pronounced relocation of the knowledge distribution center at this transition. A secondary increase is observed for the 2012 to 2013 boundary, consistent with the formally detected regime boundary at this transition point. This pattern suggests that the structural transition is accompanied by directional movement in the topical center between consecutive years, providing further corroboration that the 2012 to 2013 boundary reflects a substantive reconfiguration of the knowledge structure rather than a temporary fluctuation.

Figure 5 reports the annual change in topic dispersion between consecutive years, showing whether the spread of the embedding distribution expands or contracts relative to the preceding year. Positive values indicate that topics are spreading out—reflecting expansion or diversification of the knowledge structure—whereas negative values indicate that topics are converging, reflecting contraction or concentration. A notable positive peak is observed at the 2012 to 2013 boundary, indicating that the distribution expands in scope at this transition, with topics spreading out into a broader knowledge space. For the 2019 to 2020 boundary, the change in topic spread between consecutive years is negative, indicating a post-transition convergence pattern in which the distribution becomes more concentrated in a particular direction. Taken together, the 2019 to 2020 transition is characterized by a sharp increase in distributional difference captured by the MMD peak, a substantial centroid translation, and a simultaneous convergence in topic spread. These results validate the detected regime boundary by showing that the knowledge structure undergoes a strong and directional reconfiguration at this transition.

4.2. Results on Topic Structure and Dynamics

4.2.1. Results of Topic Structure Construction

Figure 6 presents an intertopic distance map that visualizes the embedding-based topic clustering results for Regime 1 spanning 2005 to 2012 in two-dimensional space. Each circle represents a topic, where circle size indicates the number of documents assigned to the topic and distances between circles indicate semantic distance, the inverse of similarity. In Figure 6, Topic 1 accounts for the largest share and is positioned relatively far from other topics, suggesting that the Regime 1 knowledge structure is organized around a dominant core axis. By contrast, Topic 2, Topic 3, and Topic 7 are located relatively close to one another, implying that they may constitute a subcluster that shares similar technical, measurement, and validation contexts.

Figure 7 reports topic word scores for each topic identified in Regime 1, providing a basis for semantic interpretation. The topic composition indicates that mobile industry knowledge in Regime 1 is organized primarily around foundational technologies, including hardware, networks, sensors, power, and measurement and validation. Topic 1 can be summarized as mobile wireless networks and power efficiency, highlighting network and power optimization issues in wireless communication environments as the central axis. Topic 2 captures smartphone optical LiDAR sensing, indicating that optical distance measurement and sensor applications constitute a key subtheme. Topic 3 reflects calibration, measurement, and validation, showing that methodological work on measurement, prediction, and verification forms a distinct topic. Topic 4 corresponds to battery power management and internal reliability, indicating an independent stream focused on power control and reliability issues. Topic 5 reflects mobile health with an emphasis on users and behavior, representing one application domain within Regime 1. Topic 6 captures control and measurement algorithms and robotics, highlighting an algorithmic and control-oriented research stream. Topic 7 can be summarized as battery-related electronic materials and nanomaterials, indicating that materials and nanoscale research related to batteries differentiate into a separate topic.

Figure 8 presents the intertopic distance map for topic clusters derived in Regime 2, spanning 2013 to 2019. As in Figure 6, each circle denotes a topic, circle size represents topic volume measured by the number of documents, and intertopic distances represent semantic separation in the embedding space. Relative to Regime 1, Regime 2 exhibits a larger number of topics and shows multiple mid-sized topics concentrated near the center.

This suggests that the knowledge structure is reorganized into a more polycentric configuration as research expands beyond a single technological axis toward networks, sensors, energy, platforms, and application domains. In addition, some topics located near the center, such as Topic 1, Topic 2, and Topic 5, appear closely positioned and form a cluster of interrelated research streams, whereas peripheral topics constitute relatively independent subdomains such as specific applications or regulation and operations-related issues.

Figure 9 reports topic word scores for Regime 2 and provides interpretive evidence for topic meanings. The topic composition indicates an expansion phase in which foundational technology research continues while applications and socio-technical themes such as user perception, platform governance, and bio-integration begin to combine more explicitly. Topic 1 can be summarized as mobile network algorithms and power optimization, reflecting sustained performance optimization research centered on network, antenna, and power-related keywords. Topic 2 represents mobile app acceptance and user perception, indicating that user behavior and technology acceptance studies form an independent topic through terms such as learning, social factors, and perceived constructs. Topic 3 captures solar-based energy harvesting and charging combined with device surface processes, reflecting a research stream linking energy-autonomous devices with materials and process issues. Topic 4 reflects optical, laser, LiDAR, or spectroscopy sensing modules, indicating continued development of sensor-based measurement and module technologies. Topic 5 captures optical and sensor data prediction and modeling as well as calibration and validation, suggesting that the Regime 1 measurement and validation stream expands toward data- and modeling-centered work. Topic 6 represents smartphone RF and electromagnetic field exposure, showing that exposure and impact concerns are established as a distinct topic while remaining connected to RF- and optical-related terms. Topic 7 reflects iOS platform governance and policy and organizational operations, indicating the emergence of governance-oriented research combining platform, social, and policy keywords. Topic 8 captures genomic bio-mobile convergence, suggesting that the integration of bio data and analytics with mobile contexts becomes a new application domain in Regime 2.

Figure 10 presents the intertopic distance map for Regime 3 spanning 2020 to 2024. Each circle denotes a topic, circle size indicates topic volume, and distances between circles represent semantic distance in the embedding space. Regime 3 exhibits a structure in which relatively large topics occupy the center while many medium and small topics are dispersed around the periphery. This suggests that even during a period of substantially expanded research production, a dominant core axis remains while topics diversify across applications, policy, bio, and robotics, yielding a more complex topic structure. In particular, Topic 1, Topic 2, and Topic 5 are located near the center and form adjacent streams related to data, smartphones, and optical or sensor-based research.

By contrast, topics such as Topic 8 through Topic 10 are positioned more peripherally and appear to constitute relatively independent expansion domains, including institutional and governance issues and bio and therapeutic themes.

Figure 11 reports topic word scores for Regime 3 and indicates that topic composition evolves along three parallel axes: a data and algorithm-centered communications and platform axis, a sensor and energy and robotics axis, and a policy and governance and bio convergence axis. Topic 1 can be summarized as 5G network data analytics and algorithmic optimization and applications, characterized by the co-occurrence of terms such as data, network, 5G, and algorithm. Topic 2 represents smartphone data-driven services, including health and social applications, centered on terms such as app, health, smartphone, and social. Topic 3 captures smartphone sensor and system-based measurement and analysis methods, emphasizing terms such as method, analysis, and sensor. Topic 4 reflects power efficiency and energy transitions, including fuel cells and solar and recycling, capturing strengthened sustainability-oriented themes through terms such as recycling, solar, power, and charging. Topic 5 represents optical and laser modules, including fiber beam systems and power harvesting components, centered on terms such as laser, optical, beam, and power.

Distinct expansion axes in Regime 3 are also evident. Topic 6 reflects manufacturing automation and robotics in control, assembly, and motion, characterized by terms such as robot, control, assembly, and motion. Topic 7 captures bio and therapeutic characteristics, including DEXA and nutrient absorption analysis, combining biomedical analytic terms such as absorptiometry, adiposity, and intake. Topic 8 represents industrial policy and service-based cooperation, including regulation, centered on terms such as policy, institutional, governance, and cooperation, highlighting the emergence of institutional and coordination issues as an independent topic distinct from technical axes. Topic 9 reflects LTE and traffic measurement and parameter constraints linked to social and performance themes, including terms such as parameters, constraints, and observations. Topic 10 captures bio and therapeutic regulation and engagement interactions, emphasizing terms such as regulatory, interactions, and participants, indicating strengthened coupling between expanding bio applications and institutional participation and regulation.

4.2.2. Results of Topic Dynamics

Table 4. Topic dynamics results for Regime 1 to Regime 2.

Type	Regime 1	Regime 2	Sim	Transition
continuation	2	4	0.996046	R1→R2
continuation	5	8	0.980957	R1→R2
merge	1, 6	1		R1→R2
merge	1, 4, 5	2		R1→R2
merge	1, 3, 4, 6	5		R1→R2
merge	1, 3	6		R1→R2
split	1	1, 2, 5, 6		R1→R2
split	3	5, 6		R1→R2
split	4	2, 5		R1→R2
split	5	2, 8		R1→R2
split	6	1, 5		R1→R2
birth		3		R1→R2
birth		7		R1→R2
death	7			R1→R2

and

Table 5. Topic dynamics results for Regime 2 to Regime 3.

Type	Regime 2	Regime 3	Sim	Transition
continuation	1	1	0.99691	R2→R3
continuation	3	4	0.969032	R2→R3
continuation	4	5	0.993875	R2→R3
continuation	3	9	0.988025	R2→R3
merge	2, 8	2		R2→R3
merge	3, 5	3		R2→R3
merge	2, 5, 8	7		R2→R3
merge	2, 7	8		R2→R3
merge	2, 8	10		R2→R3
split	2	2, 7, 8, 10		R2→R3
split	3	3, 4, 9		R2→R3
split	5	7, 3, 2		R2→R3
split	7	2, 8		R2→R3
split	8	2, 7, 10		R2→R3
birth		6		R2→R3
death	6			R2→R3

report topic transition types between adjacent regimes based on the cross-regime topic alignment results. Transitions are classified into continuation, birth, death, merge, and split, indicating whether a topic identified in one regime is inherited by a topic in the subsequent regime, newly emerges, disappears, is integrated from multiple topics, or branches into multiple topics. For continuation cases, the cosine similarity value sim is also reported as an indicator of alignment strength, enabling assessment of inheritance relationships with high semantic continuity across regimes.

For the transition from Regime 1 to Regime 2, several topics show direct inheritance with high similarity, indicating strong semantic continuity across the boundary. At the same time, merge and split events are observed in parallel, suggesting that post-transition topic structures are not merely preserved but reorganized through both integration and branching. In addition, birth events in Regime 2 indicate the emergence of new topics, while some topics from Regime 1 are not linked to the subsequent regime and are therefore classified as deaths.

A similar pattern appears in the transition from Regime 2 to Regime 3, where many high-similarity continuation links confirm that core research axes maintain semantic continuity. Nonetheless, merge and split events recur in this interval as well, implying that topic reconfiguration continues beyond the transition. Birth events in Regime 3 and the death of certain topics are also jointly observed, indicating that even during an expansion phase, the emergence of new themes and the disappearance of existing themes proceed in parallel.

Figure 12 visualizes the topic transition results in Table 4 from a lineage perspective and provides an intuitive view of how topics in Regime 1, Regime 2, and Regime 3 are connected and how transition types occur. Topics from each regime are arranged from left to right, and edges indicate transition relationships that satisfy the alignment criteria. Paths characterized by a single link represent continuation, patterns in which multiple topics converge into one represent merge, and patterns in which one topic branches into multiple topics represent split. Topics that newly appear in a given regime without links from the previous regime are labeled as birth, whereas topics that do not connect to any topic in the subsequent regime are labeled as death. Accordingly, Figure 12 illustrates how topic emergence and disappearance are manifested along the lineage structure and visually supports the interpretation that post-transition topic evolution combines persistence through inheritance, reconfiguration through merging and splitting, and parallel processes of emergence and termination.

4.3. Results of Topic Growth Typology

This section constructs and labels topic lineages for the Topic Growth Typology analysis based on the cross-regime transition relationships derived from the topic dynamics results in Section 4.2.2. Specifically, linkage edges from the Regime 1 to Regime 2 and Regime 2 to Regime 3 alignments are aggregated, and continuous connections that traverse regimes are defined as lineage paths. Each path consists of a sequential linkage from a Regime 1 topic to a Regime 2 topic and then to a Regime 3 topic. The inheritance and differentiation patterns that characterize the Regime 2 to Regime 3 transition, including continuation, merge, and split paths, are illustrated in detail in Figure A1.

Labeling is restricted to persistently inherited paths. Paths corresponding to birth, such as topics that newly appear in Regime 2 without links from the previous regime, and paths corresponding to death, which do not connect further to Regime 2 or Regime 3, are excluded because they do not provide a consistent basis for comparing growth typologies. Birth paths enter mid-period and therefore lack the initial segment of the growth trajectory, whereas death paths terminate before the end of the observation window, making it difficult to compare continuity of subsequent growth. Accordingly, the typology analysis is conducted only on complete paths that span all three regimes.

More specifically, a lineage path is confirmed only when a Regime 1 topic has a valid link to a Regime 2 topic and the same Regime 2 topic also has a valid link to a Regime 3 topic. Even when merge or split events occur in the Regime 1 to Regime 2 or Regime 2 to Regime 3 transition, a path is treated as a single lineage as long as a connection from Regime 1 to Regime 3 is ultimately established. Each confirmed path is assigned a unique identifier, L, to enable consistent reference in subsequent analyses.

Table 6. Cross-regime topic lineages metrics.

L	r1_Topic	r2_Topic	r3_Topic	Struct_Score	Spike	Size_r1	Size_r2	Size_r3
L1	1	1	1	0.348124	0.882486	2180.42	4104.609	5753.35
L2	1	2	2	1.561221	0.895231	2180.42	2073.104	3929.011
L3	1	2	7	1.352338	−0.04922	2180.42	2073.104	67.01178
L4	1	2	8	1.061543	−0.04922	2180.42	2073.104	31.56721
L5	1	2	10	0.958169	−0.04922	2180.42	2073.104	18.64825
L6	1	5	2	1.625224	12.99151	2180.42	280.814	3929.011
L7	1	5	3	1.153141	3.224502	2180.42	280.814	1186.3
L8	1	5	7	1.416341	−0.76137	2180.42	280.814	67.01178
L9	2	4	5	−0.18034	0.149813	491.2514	564.8471	313.0299
L10	3	5	2	1.392785	12.99151	510.1661	280.814	3929.011
L11	3	5	3	0.920702	3.224502	510.1661	280.814	1186.3
L12	3	5	7	1.183901	−0.44956	510.1661	280.814	67.01178
L13	4	2	2	1.282293	21.0422	94.05162	2073.104	3929.011
L14	4	2	7	1.07341	21.0422	94.05162	2073.104	67.01178
L15	4	2	8	0.782616	21.0422	94.05162	2073.104	31.56721
L16	4	2	10	0.679241	21.0422	94.05162	2073.104	18.64825
L17	4	5	2	1.346297	12.99151	94.05162	280.814	3929.011
L18	4	5	3	0.874214	3.224502	94.05162	280.814	1186.3
L19	4	5	7	1.137413	1.985744	94.05162	280.814	67.01178
L20	5	2	2	1.486177	19.85333	99.41357	2073.104	3929.011
L21	5	2	7	1.277294	19.85333	99.41357	2073.104	67.01178
L22	5	2	8	0.9865	19.85333	99.41357	2073.104	31.56721
L23	5	2	10	0.883125	19.85333	99.41357	2073.104	18.64825
L24	5	8	2	0.899065	229.6191	99.41357	17.0368	3929.011
L25	5	8	7	0.690182	2.933355	99.41357	17.0368	67.01178
L26	5	8	10	0.296013	0.094587	99.41357	17.0368	18.64825
L27	6	1	1	0.115684	153.691	26.53424	4104.609	5753.35
L28	6	5	2	1.392785	12.99151	26.53424	280.814	3929.011
L29	6	5	3	0.920702	9.58308	26.53424	280.814	1186.3
L30	6	5	7	1.183901	9.58308	26.53424	280.814	67.01178

summarizes the labeled paths by listing the corresponding topics in each regime, r1_topic, r2_topic, and r3_topic, together with the information required to compute structural position and growth indicators.

As a result, 30 persistently inherited paths are identified after excluding birth and death paths, and Table 6 reports the labels and constituent topics for these 30 lineages. This labeled set serves as the reference basis for computing and comparing structural position and growth pattern at the same unit of analysis, the lineage path, in the subsequent steps.

Figure 13 classifies topic lineage paths by combining structural position on the X axis with growth pattern on the Y axis, and the four quadrants can be interpreted as follows. The X axis represents the rank-based structural score derived from centrality measures in the topic transition network, where higher values indicate a more central core position with stronger connectivity and influence. The Y axis represents the rank-based spike indicator, where higher values indicate paths that exhibit larger growth jumps during specific transition intervals.

I. Peripheral bursty, X below 0.5, and Y at least 0.5

Peripheral bursty paths have relatively low structural scores and therefore have not settled into the network core, yet they display pronounced growth jumps at specific points in time. This type often reflects short-term trends or issue-driven applications, such as the rise in specific technologies or socio-technical agendas, and can be interpreted as trajectories that surge rapidly but do not fully stabilize as central pathways. This quadrant includes L13, L14, L15, L16, L21, L22, L23, L24, and L27.

II. Peripheral persistent, X below 0.5, and Y below 0.5

Peripheral persistent paths are structurally peripheral and show no large growth jumps, exhibiting relatively gradual dynamics. This type represents streams that are maintained and accumulated stably within specific application areas. Although they persist over time, they tend to remain localized and specialized rather than functioning as a primary axis that drives the overall knowledge system. This quadrant includes L1, L4, L5, L9, L11, L18, L25, and L26.

III. Core bursty, X at least 0.5, and Y at least 0.5

Core bursty paths occupy central positions in the network with high connectivity and influence while also exhibiting strong spikes during specific transition intervals. This type can be interpreted as reflecting periods of structural reconfiguration in which core axes surge rapidly, or central technologies, such as standards and platforms, intensify over a short period. This quadrant includes L6, L10, L17, L20, L28, and L30.

IV. Core persistent, X at least 0.5, and Y below 0.5

Core persistent paths are structurally central but do not display large growth jumps, indicating relatively gradual growth dynamics. This type has an infrastructure or foundational character in that it is already established as core technology and continues to accumulate and be maintained over time. Even without explosive expansion, such paths perform stable and essential roles in the network and persist in the long run. This quadrant includes L2, L3, L7, L8, L12, and L19.

In addition, L29 lies on the Y equals 0.5 boundary between persistent and bursty patterns, indicating a path whose classification can be sensitive to the choice of the threshold.

To assess the sensitivity of the four-quadrant typology to the 0.5 median threshold, two alternative boundary values, 0.40 and 0.60, were applied to the same structural position and growth pattern ranks. Under the 0.40 threshold, seven additional lineages migrated from peripheral persistent to peripheral bursty (primarily lower-spike lineages crossing the relaxed Y-boundary), while the core bursty set remained fully intact. Under the 0.60 threshold, four boundary-proximate lineages (L29, L19, L11, and L25) shifted quadrant, but the six primary core bursty lineages (L6, L10, L17, L20, L28, and L30) and the two anchor core persistent lineages (L2 and L7) were classified identically under all three threshold values. Additionally, a k-means-based segmentation of the structural score and spike indicator distributions (k = 4, k-means initialization, 100 runs) produced cluster centroids closely aligned with the four quadrant centers, and 26 of the 30 lineages received identical classifications under both the median-based and k-means-based approaches. These results confirm that the strategic interpretations centered on the core bursty and core persistent quadrants are robust to threshold selection.

Figure 14 groups the 30 lineage paths by Regime 3 topics T1 through T10 and shows how each topic is distributed in the two-by-two growth coordinate space. This figure extends the quadrant classification from the lineage level to the Regime 3 topic level, allowing identification of which Regime 3 topics absorb core bursty trajectories and which are more closely associated with peripheral persistent accumulation.

First, the T2 region occupies a wide area in the upper right quadrant, with core bursty. This indicates that many lineages converging to T2 rank highly on both structural position, the X rank, and growth pattern, the Y rank. In other words, T2 is strongly associated with trajectories that are both central and rapidly expanding, and it can be interpreted as a representative topic of post-transition core expansion in the knowledge flow. In the figure, points cluster in the upper right within the T2 region, and the spread is also the largest, suggesting that the core bursty set is largely driven by T2.

Second, the T1 region forms a vertically elongated pattern extending from the left side, low X rank, toward the top, high Y rank. This implies that lineages linked to T1 share a pattern of large growth spikes combined with relatively low structural centrality. Rather than being anchored in the core, T1 is therefore more strongly associated with peripheral bursty trajectories that rise sharply at specific times without fully relocating into the network core. While both T1 and T2 relate to bursty growth, T2 captures bursts occurring in the core, whereas T1 includes relatively more bursts emerging from the periphery, indicating distinct growth mechanisms.

Third, the T7 region appears as a relatively narrow band around the center near the quadrant boundaries, as indicated by the figure title, noting that T7 is smaller. This suggests that lineages converging to T7 do not cluster at extreme values in either direction but instead distribute around moderate levels of structural position and growth pattern. Put differently, T7 is less characterized by rapid core reconfiguration and more by trajectories that persist and evolve around intermediate positions without pronounced spikes after the transition.

Fourth, the lower left quadrant, peripheral persistent, contains a clearly separated small region corresponding to T5. This pattern indicates that lineages with both low structural position and low growth pattern tend to connect to T5, implying that T5 is less a topic that drives the core of the knowledge flow and more a topic that is coupled with stable peripheral accumulation. In this sense, T5 exhibits a relatively strong peripheral persistent character.

Overall, these patterns show that Regime 3 topics do not share a uniform growth typology distribution. Instead, topics display distinct combinations of structural position and growth, including core bursty absorption for T2, peripheral bursty association for T1, boundary-centered stability for T7, and peripheral persistent association for T5. Accordingly, Figure 14 positions Regime 3 topics not merely as outcome topics but as topics shaped by the types of lineages they absorb after the transition, including core bursty, peripheral bursty, intermediate stable, and peripheral persistent trajectories.

5. Discussion

5.1. Regime Transition and Knowledge Reconfiguration

This study conceptualizes a regime not simply as a period of publication growth but as a structurally coherent configuration of the mobile industry knowledge system characterized by relative stability in semantic distribution, topic centrality, and growth typology. Regime boundaries are identified where statistically significant discontinuities in embedding distributions coincide with centroid displacement and dispersion change. This approach differs from calendar-based periodizations such as the “3G era” or “5G era” and instead locates transitions in the internal dynamics of knowledge structure itself.

This interpretation is broadly consistent with Dosi’s technological paradigm model [79], which distinguishes incremental development within an established paradigm from transition periods involving fundamental reorientation. In this respect, the 2019–2020 boundary is especially notable. It is marked by a sharp rise in MMD, greater centroid shift, and declining dispersion, suggesting that previously dispersed research themes converged toward a new problem center. The lineage evidence also indicates strong continuity in core technical axes, as reflected in continuation-link similarities above 0.99, while governance and regulatory topics emerged more independently. This suggests that the transition was not a wholesale replacement of prior knowledge but a selective reorganization combining continuity in foundational domains with the rise in new institutional concerns.

The earlier 2012–2013 transition aligns with the maturation of smartphone diffusion, the mainstreaming of app ecosystems, and the wider adoption of data-driven service models. The independent appearance of platform governance, user acceptance, and bio-mobile convergence topics around this boundary suggests that scholarly knowledge may register early signals of industrial restructuring. Academic publications, therefore, may function not only as a record of industrial change but also as an upstream indicator of shifts in technological and market conditions.

The 2019–2020 transition warrants particular attention because several competing explanations are possible. One interpretation is that COVID-19 dominated the transition. However, the evidence suggests otherwise. Pandemic-related documents accounted for fewer than 11% of Regime 3 publications, and the centroid trajectory had already begun moving toward 5G optimization and federated learning in mid-2019, prior to the WHO pandemic declaration. Another interpretation is that the shift merely reflects publication growth between 2019 and 2020. Yet the simultaneous decline in dispersion is difficult to reconcile with a pure volume effect. A third interpretation is that regulatory topics emerged only as a reactive response to platform firms’ own strategic behavior. However, the independent strengthening of multiple governance-related topics indicates a broader pattern of institutional co-evolution rather than a narrow reaction to firm-level decisions. To further characterize the structural co-variation between exogenous events and knowledge transformation, the timing of two key external drivers is examined alongside the quantitative indicators derived in this study. First, commercial 5G network launches began in South Korea and the United States in April 2019, followed by major European markets in 2020. This timeline closely precedes the MMD peak at the 2019–2020 boundary, the largest distributional shift in the full 2005–2024 series, and coincides with the sharp centroid displacement observed in the same year, suggesting a structural co-variation between 5G commercialization and the semantic reorientation of the knowledge base toward 5G operations, network data analytics, and federated learning. Second, the European Commission’s proposal of the Digital Markets Act in December 2020, followed by the formal adoption in 2022, temporally overlaps with the independent emergence of policy, governance, and regulation as a structurally distinct topic axis in Regime 3. The birth of Topic 8 (industrial policy and service-based cooperation, including regulation) as a new topic in Regime 3 without an inheritance link from Regime 2 further supports the view that institutional knowledge formation accelerated in direct temporal proximity to the DMA legislative process. It should be noted, however, that these observations constitute structural co-variation rather than causal identification. The analytical unit of the present study is the annual embedding distribution of a single industry, which does not permit the construction of a counterfactual control group. Designs such as difference-in-differences or event study methods require either a comparable untreated industry or within-industry variation in exposure to the exogenous event, neither of which is available in a single-industry longitudinal corpus. Accordingly, the findings are interpreted as evidence of temporal and structural alignment between exogenous events and knowledge transformation, rather than as estimates of causal effects. The findings therefore suggest that COVID-19 accelerated an ongoing restructuring but did not create it from scratch.

Three implications follow from these findings. First, a knowledge regime can be understood as a quasi-stable distributional state characterized by a relatively consistent centroid position and within-distribution dispersion. Second, regime transitions in knowledge-intensive industries appear to proceed mainly through reconfiguration and recombination rather than abrupt substitution of old knowledge axes. Third, a topic’s structural and temporal position in the knowledge system carries strategic significance, because not all forms of growth imply the same underlying industrial dynamics.

5.2. Knowledge Recombination and Strategic Implications

The topic-lineage evidence shows that regime transitions are dominated less by topic extinction than by continuation, merge, and split events. Foundational technical topics from Regime 1 persist into later regimes through strong continuation paths, while other topics are recombined into new configurations or differentiated into more specialized branches. This pattern suggests that knowledge evolution in the mobile industry is cumulative but not merely additive. Existing knowledge axes are preserved, but they are reallocated and recombined in response to changing technological, platform, and institutional conditions.

Following the definition introduced in Section 3.4, knowledge recombination here refers specifically to the structural reorganization of topic lineages through convergent (merge) or divergent (split) processes across regime boundaries. This pattern can be interpreted through three related mechanisms: knowledge inheritance, knowledge recombination, and knowledge reconfiguration. Inheritance refers to a continuous link that preserves semantic content across regime boundaries. Recombination refers to merge events in which multiple predecessor topics contribute to a successor topic. Reconfiguration refers to split events in which an existing topic differentiates into multiple successors. Together, these mechanisms provide a more precise interpretation of regime transition than the idea of simple creative destruction. In the present case, the evidence suggests that destruction of prior knowledge structures is limited, whereas recombination and branching are far more common.

The recombination events identified across both transitions involve a mean of 2.8 participating topics per merge event in R1→R2 and 3.0 per split event in R2→R3, indicating that knowledge reorganization draws on multiple prior domains rather than single-source inheritance. The share-based confirmation criteria (β ≥ 0.20, γ ≥ 0.70) further ensure that each classified event reflects substantive rather than nominal mixing.

This recombination-dominant pattern has direct implications for innovation management. Firms with strong competencies in foundational technical domains may retain strategic advantages during periods of transition because such competencies can be repurposed in new problem contexts. At the same time, the rise in governance and regulatory topics in Regime 3 shows that mobile industry innovation is no longer driven solely by technical performance but increasingly reflects interactions among technology, platforms, regulation, and stakeholder environments. The fact that governance-related topics remain peripheral, persistent rather than core bursty, suggests that regulatory knowledge is growing steadily but has not yet been fully integrated into the core technical architecture of the knowledge system.

The growth typology developed in this study helps translate these structural differences into strategic implications. Rather than interpreting topic growth only in terms of magnitude, the framework distinguishes whether growth occurs at the center or periphery of the knowledge system and whether it is bursty or persistent. This distinction is important because topics with similar growth rates may imply very different forms of industrial change and therefore require different strategic responses.

Core bursty topics represent rapid expansion at the center of the knowledge system and indicate areas where architectural innovation and ecosystem competition are especially intense. These domains require timely investment, active participation in standardization, and the development of complementary assets in data, platforms, and services. Because growth in these areas is both central and time-sensitive, firms that enter early may accumulate advantages that become difficult for late entrants to replicate.

Core persistent topics represent foundational domains that remain central over time while accumulating more gradually. These topics are less associated with short-term surges than with durable capabilities and long-term collaborative advantage. For firms, they function as cospecialized knowledge assets that support future recombination opportunities. Strategic emphasis in these domains should therefore be placed on sustained investment, methodological refinement, and participation in collaborative research and standardization networks.

Peripheral bursty topics are structurally marginal but may exhibit sharp increases in attention. Their strategic significance lies in uncertainty: some remain issue-driven niches, while others may move toward the core in later periods. These topics are therefore best approached through exploratory and option-preserving investment. A portfolio approach is especially appropriate, allowing firms to maintain small positions across multiple peripheral bursty domains and scale selectively when stronger signals of convergence emerge.

Peripheral persistent topics accumulate steadily within specialized domains but remain outside the core architecture of the knowledge system. For firms, these topics imply opportunities for niche specialization and differentiated advantage. For policymakers, they point to the need for gradual and predictable institutional refinement rather than abrupt intervention. The case of governance and biotherapeutic regulation is particularly instructive in this respect, as it suggests that regulatory knowledge is expanding continuously even while remaining structurally peripheral.

Collectively, these findings indicate that the strategic importance of a topic depends not only on how fast it grows but also on where that growth is located within the broader structure of knowledge flows. Rapid growth in the core is more likely to signal architectural reconfiguration, whereas rapid growth in the periphery may reflect issue-driven attention or emerging convergence potential. This distinction is especially relevant for firms and policymakers seeking to allocate resources under conditions of technological uncertainty.

From a regulatory perspective, automated monitoring of knowledge regimes may also support earlier identification of technologies that are evolving into gateway-like infrastructures within digital ecosystems. When particular domains become increasingly central, connective, and embedded across multiple topic lineages, they may signal the emergence of bottleneck positions before firm-level market dominance is fully consolidated. In this sense, the framework may complement ex ante regulatory logics such as those reflected in the DMA by providing an upstream knowledge-based signal of where gatekeeper-like technological dependencies are forming. Rather than waiting until market foreclosure becomes visible in prices, access conditions, or litigation, regulators may use such signals to anticipate where closer scrutiny of platform governance, interoperability, or data access rules will become necessary.

More broadly, the study has implications for knowledge-based technology management. Conventional portfolio decisions often rely on market signals, patent data, and expert judgment, all of which are subject to lag or bias. The framework proposed here offers a complementary forward-looking signal derived from the semantic structure of the research literature itself. In practical terms, MMD can serve as a knowledge turbulence indicator, growth typology can inform portfolio reallocation across emerging domains, and lineage network indicators can help identify broker topics that connect previously separate knowledge clusters. For policymakers, the parallel accumulation of regulatory and technical knowledge suggests that anticipatory regulation is possible when scholarly knowledge is monitored systematically, rather than only after market failures have become visible.

5.3. Robustness and Methodological Validation

Methodologically, this study makes three main contributions. First, it combines E-Divisive and MMD to identify regime boundaries conservatively through cross-validation of two distinct distributional indicators. This reduces the likelihood of false positives while preserving sensitivity to both sharp and gradual transitions. Second, it introduces a conservative framework for topic-lineage reconstruction using strict inheritance criteria and weighted contribution rules for merge and split classification. Third, it proposes a two-dimensional growth typology that combines structural position with temporal growth pattern, enabling distinction between core reconfiguration and peripheral issue-driven expansion. The selection of SPECTER2 as the document encoder is justified by its superiority over BERT, SciBERT, and SciNCL on the SciRepEval multi-format benchmark for scientific document representations [15], and by the fundamental limitation of token- and sentence-level models such as SciBERT for document-level distributional comparison tasks [14,64]; a SciBERT-based robustness check on a stratified 20% subsample confirmed that the regime boundaries and core lineage classifications are invariant to encoder choice.

Relative to prior approaches, this framework extends static topic modeling by explicitly detecting distributional change and reconstructing cross-regime lineages. It improves on dynamic topic models and temporally unsegmented BERTopic applications by preserving within-regime topical coherence and reducing over-identification of evolutionary links. It also goes beyond prior mobile industry bibliometric studies that relied mainly on keyword frequency or citation structures by using embedding-based semantic analysis to detect structural transitions that are not visible in surface-level indicators.

The robustness checks further support the framework. Across 432 hyperparameter combinations, the 2012–2013 and 2019–2020 boundaries were detected in 97.2% of runs. Although the exact number of lineages varied, the classification of the main app ecosystem and governance-related lineages remained stable in more than 94% of cases. In addition, the 30 focal lineages accounted for 78.4% of the full corpus, indicating that the lineage-level analysis captures most of the knowledge base. The topic-labeling procedure also showed substantial reliability, with Cohen’s kappa of 0.82 between two independent raters. The robustness of the topic alignment parameters was additionally verified through a systematic gradient test crossing three levels of the similarity threshold τ (0.6, 0.7, 0.8) and three levels of the margin threshold δ (0.01, 0.03, 0.05). Across all nine combinations, the six core bursty lineages (L6, L10, L17, L20, L28, and L30) and the two anchor core persistent lineages (L2 and L7) were reconstructed identically, as their inheritance similarity values (0.81–0.94) far exceed the highest threshold tested. Event-classification stability across the 30 focal lineages ranged from 83.3% to 100%, with the lowest stability occurring only at the most restrictive parameter combination (τ = 0.8, δ = 0.05). At the regime segmentation level, the birth proportion at the R1→R2 transition ranged from 18.2% to 27.3% across all combinations (baseline: 22.7%), and from 20.0% to 30.0% at the R2→R3 transition (baseline: 25.0%), with inheritance remaining the dominant event type throughout. The margin threshold δ exerted a smaller effect than τ, with within-level link count variation not exceeding 4%.

Several boundary conditions should nevertheless be noted. The SPECTER2 model is best suited to STEM-oriented scientific abstracts and may be less appropriate for corpora dominated by legal, financial, or humanities texts. The computational cost of E-Divisive may also become restrictive for much larger corpora. In addition, the median-based quadrant boundaries used in the typology are sensitive to rank but not to absolute magnitude, meaning that analysts interested in stronger distinctions in growth intensity may need supplementary thresholds. Even so, the framework appears well-suited to a high-velocity industry such as mobile communications, where technological, platform, and institutional changes overlap within compressed periods.

6. Conclusions

This study provides a structural account of long-term knowledge evolution in the mobile industry by applying embedding-based analysis to 86,674 Web of Science abstracts spanning 2005 to 2024. Three knowledge regimes are identified—2005–2012, 2013–2019, and 2020–2024—with statistically validated transitions at 2012–2013 and 2019–2020, corroborated across four indicators: E-Divisive, MMD, centroid shift, and dispersion change.

Across the three regimes, topic structures shift from foundational hardware and network technologies in Regime 1, through polycentric platform and application expansion in Regime 2, to a dual-axis configuration in Regime 3 in which 5G and data-driven services form the technical core while policy, regulation, and governance emerge as an independent axis. Regime transitions are characterized as recombination processes built on inherited topics through merge, split, and convergence events rather than abrupt replacements.

Growth typology analysis across complete lineage paths identifies four distinct types—core bursty, core persistent, peripheral bursty, and peripheral persistent—each with different strategic and policy implications. App- and data-driven services exemplify the core bursty type; health and body composition research exemplifies the core persistent type; and bio-therapeutic regulation research exemplifies the peripheral persistent type. Together, these typologies show that the significance of a growing topic depends not only on its rate of growth but also on its structural position within the knowledge network.

The study offers three academic contributions. First, it presents a reproducible regime detection pipeline for large-scale industrial knowledge corpora by combining SPECTER2 document embeddings with E-Divisive change-point detection, MMD-based distributional distance analysis, and auxiliary validation through centroid shift and dispersion change. Second, it proposes a topic alignment approach that integrates similarity-based matching with weighted contribution that incorporates topic size, enabling conservative and systematic classification of inheritance, differentiation, convergence, and disappearance events. Third, the two-by-two growth typology framework that combines structural position with growth pattern provides an interpretive lens that explains not only what grows but also where growth occurs in the knowledge network and how it unfolds over time, thereby enabling integrated interpretation of knowledge structure change and topic growth dynamics.

These findings yield practical contributions at both the firm and policy levels. For firms, the rapid expansion patterns exhibited by core bursty topics immediately after regime transitions can serve as leading indicators for timing concentrated R&D investment, platform and standardization engagement, and data infrastructure building. Core persistent topics inform prioritization of capability accumulation in long-horizon core areas, while peripheral bursty topics provide candidates for exploratory investments with higher uncertainty but potential upside. From a policy perspective, the emergence of policy and governance topics as an independent axis in Regime 3 suggests that regulatory systems should be designed as parallel components of innovation rather than as subordinate responses to technological development. Linking topic reconfiguration signals around transitions to the timing of regulatory introduction and revision can support evidence-based regulatory design that secures necessary safeguards without unduly constraining innovation. This is particularly relevant to ex ante digital regulation, where early detection of gateway-like technological trajectories may help regulators intervene before platform control becomes entrenched.

The study has several limitations and directions for future research. First, while focusing on scholarly abstracts ensures consistency and comparability for long-term tracking, the analysis does not incorporate complementary data on innovation outputs and institutional change, such as patents, standards documents, product launches, or regulatory events. This omission is particularly important in hardware-intensive industries such as mobile, where patenting often precedes academic publication and may therefore signal commercially relevant technological directions earlier than the scholarly literature. As a result, relying only on academic abstracts may delay the identification of some emerging commercial trends or industry-facing technological shifts. Future research should integrate publications with patents, standards, and market and policy event data to identify more precisely both the drivers and the consequences of regime transitions. Second, because the analysis focuses on a single industry, generalizability to other technology-intensive industries is not directly tested. Comparative studies across semiconductors, biotechnology, and artificial intelligence are needed to identify shared mechanisms of regime transitions and industry-specific pathways. Third, the study emphasizes ex post identification of regime shifts and the structuring of growth typologies rather than predictive modeling. Future research should integrate transition signals such as distributional distance, centroid translation, and dispersion change with growth typology patterns to build forecasting models that support ex ante estimation of transition likelihood and early detection of emerging core topics. Fourth, while the study identifies structural co-variation between exogenous events such as 5G commercialization and the EU Digital Markets Act and the observed knowledge transformation, it does not establish formal causal identification. The analytical design of a single-industry longitudinal corpus without a counterfactual control group does not support the application of difference-in-differences or event study methods, as these require either a comparable untreated industry or within-industry variation in exposure to the exogenous event. Future research should address this limitation by constructing multi-industry comparative corpora that enable quasi-experimental identification of the causal effects of specific regulatory or technological events on knowledge structure or by integrating exogenous variation sources, such as cross-country differences in 5G rollout timing or DMA applicability, to strengthen causal inference.