Technology Analysis of Extended Reality Using Machine Learning and Statistical Models

Abstract

Extended reality (XR), encompassing augmented reality (AR), virtual reality (VR), and mixed reality (MR), is a key enabling technology for virtual worlds, and XR-related patents continue to grow rapidly. However, patent-based XR technology analysis faces a fundamental challenge: document–keyword matrix (DKM) built from patent titles and abstracts are typically high dimensional, sparse, and often exhibit excess zeros, which can distort inference when conventional text mining pipelines are applied without a generative count perspective. In this study, we propose a statistically grounded XR technology analysis framework that combines likelihood-based count modeling with interpretable structure mining to map XR sub-technologies from a patent DKM. Using an XR patent–keyword matrix, we fit Poisson regression (PR), negative binomial regression (NBR), and zero-inflated negative binomial regression (ZINBR) models via maximum likelihood estimation (MLE), controlling for document-length effects. Model selection by Akaike information criterion (AIC) consistently favored NBR for both target keywords, indicating substantial overdispersion in XR patent counts. We interpret exponentiated coefficients as incidence rate ratios (IRRs) and construct a technology relatedness network from significant IRR edges, revealing a dual-axis XR structure: reality is anchored in an AR or VR experience and content axis such as virtual and augment, whereas extend is embedded in a structure and integration axis for example, surface, edge, layer, and connectivity-related terms. To show how the proposed method can be applied to real domains, we searched the XR patent documents, and analyzed them for XR technology analysis.

1. Introduction

Extended reality (XR), an umbrella term encompassing augmented reality (AR), virtual reality (VR), and mixed reality (MR), has emerged as a core enabling technology for the realization of virtual worlds and immersive human–computer interaction paradigms. XR systems integrate sensing, rendering, interaction, and spatial computing components to deliver persistent and context-aware experiences that bridge physical and digital environments [1,2]. As XR hardware and software ecosystems mature, XR technologies are increasingly embedded in diverse application domains, including industrial training, manufacturing, healthcare, education, entertainment, and remote collaboration. Consequently, understanding the evolving structure of XR technologies and identifying meaningful XR sub-technologies have become important for both research and practice [1,2]. A particularly informative lens for studying technology evolution is patent data, which often capture early stage inventive activity and provide structured signals such as assignees, filing dates, classifications, and claims that can be leveraged for technology intelligence and R&D strategy [3]. In recent years, patent analytics has expanded from descriptive statistics to data-driven discovery of technology clusters, novelty, and trajectories, enabling more systematic technology landscape mapping and evidence-based decision making [4,5].

However, despite the growing strategic importance of XR, the patent-based XR technology landscape remains challenging to analyze at scale because the underlying textual evidence must be translated into machine-readable representations that support robust modeling and interpretation. A common pipeline for patent text analytics transforms patent documents into a document–keyword matrix (DKM), where rows correspond to patent documents, columns correspond to extracted keywords, and entries represent keyword frequencies [6,7,8,9]. While DKM representations enable classical machine learning workflows and facilitate interpretability, the DKM related to XR patents is typically high-dimensional and extremely sparse. In such settings, the observed number of zeros greatly exceeds what standard count models would predict, exhibiting a zero-inflated (ZI) pattern that reflects both structural absence of a technology concept in a patent and stochastic absence due to limited disclosure length, vocabulary variability, and extraction noise [10,11,12]. This sparse, zero-inflated count structure can degrade the validity of downstream analyses if ignored, leading to unstable topic discovery, biased similarity estimation, and overly fragmented clustering outcomes [10,11,12]. Much of the prior patent analytics literature has relied on probabilistic topic modeling or embedding-based clustering for technology topic identification and trend analysis, often combined with visualization or network-based mapping. These approaches have demonstrated practical utility across domains, including technology forecasting and emerging technology detection [13,14,15]. Nevertheless, many widely used workflows implicitly treat the DKM as if zeros were generated solely by a standard count process, or they apply ad hoc preprocessing such as term frequency-inverse document frequency (TF-IDF) transformations, aggressive filtering, or binarization to mitigate sparsity [16,17]. Such practices can be problematic for XR patents because the mechanism generating excessive zeros is not explicitly modeled, and the resulting sub-technology structures may be sensitive to preprocessing choices. This motivates the need for modeling strategies that are likelihood-aware for zero-inflated counts while retaining interpretability for technology mapping [18]. To address these challenges, this study focuses on XR patent DKM and proposes an analysis framework that treats extracted keywords as observable variables representing XR sub-technologies, while explicitly accounting for sparse, zero-inflated count generation. In particular, we position XR sub-technology discovery as a problem of learning low-dimensional latent structure from a high-dimensional sparse count matrix, where both structural zeros and sampling zeros may coexist [18,19,20]. Building on advances in count-based latent factor models and scalable inference for sparse matrices, we aim to extract interpretable sub-technology components and quantify their relationships in a way that is statistically principled and empirically robust [20]. In addition to sub-technology extraction, XR technology planning often requires understanding how technological elements combine and co-evolve. Therefore, we further analyze the association structure among keywords using a co-occurrence network perspective and exploit modern community detection methods to summarize dense interaction patterns into coherent technology modules. This network view complements count-likelihood modeling by highlighting technology combinations that frequently appear together, providing an intuitive map of sub-technology proximity and modular organization [21,22].

The main contributions of this work are summarized as follows. First, we introduce a statistically grounded framework for XR technology analysis by explicitly modeling patent-derived DKMs as zero-inflated count data, addressing a critical limitation of conventional text mining approaches that ignore excess zeros and overdispersion. Second, we develop a likelihood-based modeling pipeline that enables interpretable extraction of XR sub-technologies through incidence rate ratio (IRR) analysis, providing statistically meaningful measures of technology relatedness. Third, we construct a technology relatedness network based on statistically significant IRR effects, offering an intuitive and structurally grounded representation of XR technology relationships. Finally, by integrating statistical modeling with network analysis, we provide a unified and reproducible framework that improves robustness, interpretability, and practical applicability for XR patent analytics.

The remainder of this paper is organized as follows. Section 2 introduces the research background related to technology analysis and zero-inflated count modeling. Section 3 presents the data and proposed method for XR technology analysis using machine learning and statistical models. Section 4 shows the experimental results to verify the performance of our method. Lastly, we represent the conclusions of our paper in Section 5.

2. Research Background

2.1. Technology Analysis

XR, encompassing AR, VR, and MR, is widely regarded as a foundational set of technologies for immersive digital environments and virtual worlds. From a technology management perspective, XR is not a single monolithic technology but a convergent ecosystem that integrates heterogeneous components such as sensing, tracking, rendering, display hardware, interaction techniques, and enabling software stacks [23]. Because these components evolve at different rates and are driven by different industrial actors, a systematic technology analysis is essential to map the XR technology landscape, define interpretable sub-technology areas, and compare their relative maturity and relationships using reproducible evidence [23]. In this study, we focus on patent documents as a structured and scalable source to operationalize XR sub-technology variables via patent keywords and their empirical usage patterns. In general, technology analysis refers to a family of approaches used to understand the structure, dynamics, and trajectories of a target technology domain for purposes such as strategic planning, R&D portfolio management, and policy design. Existing approaches can be broadly grouped into qualitative methods and quantitative methods. Qualitative methods include expert panel review, Delphi-style elicitation, and technology roadmapping (TRM), which emphasize interpretability and contextual reasoning but can be limited by scalability and reproducibility [24]. TRM, in particular, provides a systematic way to align markets, products, and technologies over time, and recent review work has highlighted how roadmapping has evolved into more data-supported and model-driven practices across multiple industries. Quantitative methods, by contrast, rely on observable artifacts such as publications, patents, and product data, and apply bibliometrics, scientometrics, network analysis, and machine learning to generate evidence-based maps and indicators of technological change. In practice, high-quality technology analysis often combines both paradigms: qualitative judgment is used to validate and label patterns discovered from quantitative data, while quantitative analysis provides coverage, objectivity, and auditability [24]. Among quantitative data sources, patent documents are particularly valuable for technology analysis because they provide standardized metadata such as filing dates, assignees, inventors, jurisdictions, patent classification codes, and rich technical descriptions that may appear earlier than journal publications or commercial products [25,26]. A common framework to synthesize such evidence is patent landscape analysis (PLA), which aims to produce a structured overview of the development of a technology field and to identify trends, leading actors, and underexplored opportunities [25]. International guidance for preparing PLA reports emphasizes careful definition of the analysis scope, transparent search strategies, and robust data cleaning and reporting practices. More recent methodological discussions further underline that PLA can support decision-making for diverse stakeholders by connecting patent-based signals to technology trend identification and research funding or innovation policy needs [25,26].

A typical patent-based technology analysis pipeline consists of domain scoping and query design, patent retrieval and deduplication, text preprocessing and representation construction, and downstream modeling for mapping and inference [27,28]. First, the analyst defines the target domain and designs a retrieval strategy using combinations of keywords, patent classification codes, and inventor constraints, often iterating through pilot searches to improve precision and recall. Second, the retrieved corpus is cleaned by handling patent families, removing duplicates, harmonizing assignee names, and optionally restricting to specific jurisdictions or time windows depending on the research objective. Third, the textual content such as title, abstract, and claims is transformed into an analysis-ready representation through preprocessing steps such as tokenization, normalization, stopword removal, and domain-specific phrase handling. Open source guidelines and handbooks for patent analytics provide detailed, practical recommendations for implementing these steps in a reproducible manner, including common pitfalls in patent text processing and suggestions for toolchains that support scalable analysis [27,28]. To operationalize sub-technology structures from patent text, a widely used representation is the DKM, where rows correspond to patent documents, columns correspond to extracted keywords, and entries represent observed counts [13,14,29]. The DKM enables a range of downstream analyses, clustering and map-based visualization to identify coherent technology groups, topic modeling to extract latent themes that can be interpreted as technology topics, and time-aware analyses to track the emergence and diffusion of topics or clusters [13,14,29]. Prior patent analytics studies have demonstrated that text mining and clustering can help identify emerging and vacant technology areas, while topic modeling combined with expert judgment can produce interpretable technology themes and connect them to subsequent statistical analyses such as forecasting or outcome modeling. Such pipelines are directly relevant to XR patents, where the goal is to extract meaningful sub-technology variables from large-scale patent corpora in a way that supports both interpretability and empirical validation [13,14,29]. However, a key methodological complication is that patent DKM is typically high-dimensional and extremely sparse, implying that a large fraction of entries are zeros. In XR-related corpora, sparsity can be exacerbated by rapidly evolving terminology, heterogeneous subdomains, and long-tailed keyword distributions. Importantly, the abundance of zeros is not merely a computational nuisance; rather, it may reflect distinct data-generating mechanisms. Some zeros are structural, meaning that the corresponding keywords are genuinely irrelevant to a given document, whereas others are sampling zeros, meaning that the keywords are relevant but do not appear in the selected text fields. This motivates modeling strategies that go beyond standard continuous embeddings or naive frequency-based representations [30,31]. For example, prior work has introduced explicit zero-inflated formulations for learning representations from sparse count matrices, demonstrating that accounting for excess zeros can improve the faithfulness of learned structures. Likewise, the broader patent text mining literature has increasingly emphasized the importance of selecting representations and models that match the statistical properties of patent data, including sparsity and count-valued observations. Building on this perspective, the next Section 2.2 reviews zero-inflated count modeling as a principled foundation for analyzing sparse patent DKM and for defining XR sub-technology variables in a statistically coherent manner [30,31].

Recent studies have increasingly incorporated artificial intelligence and deep learning techniques into patent analytics to improve technology mapping and forecasting [31,32,33,34,35,36]. In particular, transformer-based language models and embedding-based approaches have been applied to capture semantic relationships in patent texts, enabling more accurate identification of emerging technologies and innovation patterns [37,38]. In the XR domain, recent works have focused on integrating immersive technologies with industrial applications and analyzing their technological evolution using data analysis methods. At the same time, graph-based approaches have gained attention for modeling complex relationships among technological elements, such as cooccurrence networks, citation networks, and knowledge graphs [39,40,41,42]. These methods provide complementary perspectives to traditional clustering and topic modeling by emphasizing structural relationships among technologies. However, many of these approaches do not explicitly consider the statistical properties of sparse count data, particularly zero inflation and overdispersion, which are critical in the DKM of patent data. Globally, XR research and innovation have been actively pursued across multiple regions, including North America, Europe, and East Asia. In particular, the United States, China, and European countries have played leading roles in XR technology development, with increasing integration of immersive technologies into industrial and consumer applications. These global trends highlight the relevance of XR technology analysis from an international perspective.

2.2. Zero-Inflated Count Modeling

Patent-derived DKMs are count-valued and typically extremely sparse, so a statistically principled analysis starts by treating each entry as an outcome of a count-generating process. Standard Poisson models provide a baseline but often fail for text-derived counts because the conditional variance frequently exceeds the mean, producing overdispersion. The Negative Binomial (NB) model accommodates such dispersion and is therefore commonly used as a more flexible baseline for patent and text count data [43]. In many DKM data, however, the observed proportion of zeros is substantially larger than what Poisson or NB would predict, motivating ZI models. They represent the data as a mixture of an inflation component that generates a point mass at zero and a baseline count component that generates nonnegative counts. Under this formulation, zeros can arise either from an inflation mechanism, which corresponds to structural zeros, or from the baseline count distribution, which corresponds to sampling zeros. Regression extensions link the inflation probability (e.g., via a logit link) and the baseline mean (e.g., via a log link) to covariates, yielding canonical models such as ZIP and ZINB. This decomposition aligns naturally with patent DKM, where a keyword may be absent because it is genuinely irrelevant to a document’s technical content or because it is not observed in the selected text fields despite conceptual relevance [44].

Recent advances in zero-inflated modeling and sparse count data analysis have further emphasized the importance of accounting for excess zeros in high dimensional data. In particular, zero-inflated models have been extended to large scale applications, including text mining, recommender systems, and representation learning for sparse matrices [20,30,31,44,45]. These developments highlight the necessity of adopting statistically principled approaches when analyzing the patent count data, where ignoring zero inflation may lead to biased inference and unstable model performance. A closely related class is the hurdle model, which separates the probability of observing any positive count from the positive-count magnitude through a zero-truncated count distribution. Compared with hurdle models, ZI models allow zeros to arise from both components, which can be preferable when the “at-risk” group can still produce zeros through the baseline count process. Practical guidance in the recent literature emphasizes that choosing between ZI and hurdle formulations should be driven by interpretability, diagnostic checks, and out-of-sample fit rather than by convention [46]. ZI and hurdle models are commonly fit via maximum likelihood estimation (MLE) and are supported by modern software. Likelihood-based implementations such as generalized linear mixed model (GLMM) offer efficient estimation of ZI count models within a GLMM framework, while Bayesian multilevel modeling provides flexible specification and uncertainty quantification when hierarchical structure is required. These tools facilitate transparent model comparison and reproducible reporting in applied technology analytics [47,48,49].

Beyond independent regression, ZI modeling has expanded to accommodate complex dependence structures, including multilevel, temporal, and spatial processes, and recent reviews consolidate modeling and computational strategies for such settings. These extensions are relevant for patent analytics when observations are grouped by assignee, jurisdiction, or time, or when correlation is induced by technology classes [45]. A distinctive challenge in patent text mining is the scale of the DKM. Because the matrix is dominated by zeros, scalable approaches often rely on sparse-matrix computations and inference procedures that scale with the number of nonzero entries. Representative developments include Hierarchical Poisson Factorization (HPF) for large sparse count matrices and explicitly zero-aware representation learning such as Zero-Inflated Exponential Family Embeddings, which treat zeros as statistically meaningful outcomes rather than missing information. These lines of work motivate likelihood-consistent methods for extracting interpretable latent structure from sparse patent DKM [20,30]. Taken together, the preceding discussion supports a central premise of this study: XR patent DKM are better viewed as realizations from a sparse, heterogeneous, and often zero-inflated count process, and ignoring this feature constitutes likelihood misspecification that can distort inferred technology structure. This motivates the methodology presented in Section 3, where we adopt a likelihood-based framework for zero-inflated count structure learning and complement it with network-based analysis for characterizing technology associations.

Despite these advances, existing studies on XR technology analysis and patent analytics have largely relied on frequency data analysis methods, topic modeling, or embedding clustering without explicitly addressing the zero-inflated and overdispersed nature of patent DKM. This gap motivates the need for a unified framework that integrates statistically grounded count modeling with interpretable structure analysis, as proposed in this study. The choice of statistical models in this study is guided by the intrinsic characteristics of DKM derived from patent documents. Because the data consist of sparse count observations with substantial variability, Poisson models serve as a natural baseline but may be inadequate due to their restrictive mean and variance assumption. The negative binomial model is therefore adopted to account for overdispersion commonly observed in text-derived count data. Furthermore, zero-inflated models are considered to explicitly capture excess zeros arising from both structural absence and sampling variability in patent documents.

3. Data and Proposed Method

3.1. DKM Construction and Data Desscription

To perform XR technology analysis using machine learning and statistical models, we collect the patent documents related to XR technology from the patent databases such as the United States Patent and Trademark Office (USPTO) and Korea Intellectual Property Rights Information Service (KIPRIS) [50,51]. The selection of USPTO and KIPRIS databases is motivated by their accessibility, data quality, and representativeness. The USPTO reflects a major global innovation ecosystem with extensive patent activity in XR technologies, while KIPRIS provides high-quality data from a technologically advanced and innovation-driven economy. Together, these sources offer a balanced and informative dataset for XR technology analysis. Using text mining techniques, we preprocess the collected the patent documents related to XR technology and build the DKM. After applying data cleaning, filtering, and preprocessing procedures, a total of 2500 patent documents were selected for analysis. These patents cover a wide range of XR-related technology fields, including AR, VR, and MR, as well as supporting technologies such as sensing and tracking, rendering and visualization, user interaction, display systems, spatial mapping, and communication or connectivity components. This diversity of technology domains ensures that the constructed DKM captures a comprehensive representation of XR technology components and their interactions. Algorithm 1 describes the entire process of building the DKM in this paper.

Algorithm 1 Construction of Patent Document–Keyword Matrix (DKM)

Input:  Patent records, $D=\left(d_1,d_2,\dots ,d_n\right)$ with fields of $\left({patent}_{id},title, abstract\right)$ n: number of patent documents Output:  Document–keyword matrix (DKM), $M\in N_0^{n\times p}$ p: number of patent keywords $\hspace{1em} N_0=N\cup \left\{0\right\}=\left\{\mathrm{0,1},2,\dots \right\}$ Procedure: Patent retrieval and corpus definition (1-1) Retrieve patent documents relevant to target technology using XR keyword queries (1-2) Define analysis unit by each patent document Construction of text database (2-1) Remove duplicates and consolidate patent members (2-2) Concatenate fields to construct text database $X_d$, each document d    $X_d\leftarrow {title}_d\left|\right|{abstract}_d$ Text normalization and tokenization (3-1) Apply lowercasing, punctuation, whitespace to $X_d\to {\tilde{X}}_d$ (3-2) Convert ${\tilde{X}}_d$ into tokens, apply stopword removal, lemmatization, and stemming Term generation (4-1) Generate unigram candidates and keep their within-document occurrence counts (4-2) Extract keywords from all candidate terms by computing document frequency DKM construction For each document d and each keyword k, Set $m_{dk}\leftarrow$ count of keyword k in document d $M=\left(m_{dk}\right)$ is DKM Compute sparsity, zero ratio (ZR), $ZR={\displaystyle \frac{1}{DK}}{\sum }_{d=1}^D{\sum }_{k=1}^{K}1\left(m_{dk}=0\right)$

We construct the DKM from XR patent documents with their titles and abstracts using Algorithm 1. Our DKM has a structure as shown in the following Figure 1.

So, our study analyzes the patent-derived DKM constructed from patent documents. Each row corresponds to a patent document and each column corresponds to a keyword extracted from the patent title and abstract. The matrix entry $x_{dk}\in N_0$ denotes the frequency count of keyword v observed in document d. Because the DKM is typically sparse, the matrix exhibits a high proportion of zeros, which is consistent with zero inflation and motivates count models that explicitly account for excess zeros. Prior to modeling, we perform two data cleaning steps. First, documents with all-zero keyword counts are removed, since they contain no usable information for keyword-based modeling. Second, if duplicated keyword columns occur due to preprocessing or name repair during data export, for example duplicated columns indexed with suffixes, we consolidate them by summing counts across identical keyword labels. After these steps, the cleaned DKM is used for all subsequent analyses. To control for document length effects, we compute the document level total count $L_d=\sum _k{m}_{dk}$ and define a length covariate ${loglen}_d=log\left(L_d+1\right)$. This covariate serves as a normalization control because longer texts tend to produce larger counts for many keywords, which can confound the estimation of keyword-specific associations. Since $L_d$ generally tends to be skewed towards one side, we take the log here to smooth out the extreme values. Also, to prevent $log\left(0\right)$ when $L_d=0$, we add 1 to $L_d$. To ensure reproducibility, the preprocessing steps were implemented using standard text mining procedures, including tokenization, stopword removal, lemmatization, and stemming. Keywords were selected based on document frequency thresholds, and only terms appearing in at least 20 documents and less than 90% of all documents were retained. This filtering step reduces noise from extremely rare or overly common terms. In this study, the analysis is based on an aggregated DKM, and temporal information such as patent filing year is not explicitly incorporated. This design allows us to focus on the overall structural relationships among XR technologies.

3.2. Pipeline of XR Technology Analysis

To address these challenges, we propose a structured analysis pipeline for XR technology. The pipeline consists of three main components. First, we apply likelihood-based count models to capture the statistical properties of sparse and zero-inflated data. Second, we estimate conditional relationships among keywords using interpretable regression coefficients. Third, we construct a technology relatedness network based on statistically significant associations, enabling intuitive visualization of XR sub-technology structures. The pipeline consists of four components. First, we carry out the zero-inflated count modeling for target technologies using Poisson and negative binomial (NB) distributions and zero-inflated models based on Poisson and NB distributions. Second, we perform the technology relatedness network construction using statistically significant keyword effects. Finally, we explore the white space and prediction analyses for vacant technology forecasting, including candidate keyword-pair opportunities and document-level high predicted but low observed cases.

The selection of these models is motivated by the statistical properties of the DKM. In particular, keyword frequencies are non-negative integer counts, making count regression models appropriate. However, due to the high sparsity and heterogeneity of patent text data, simple Poisson assumptions are often violated. Therefore, we extend the modeling framework to include negative binomial and zero-inflated variants, which provide greater flexibility in capturing overdispersion and excess zeros. In this study, the response variable represents the frequency count of a selected target keyword in each patent document. Predictor variables are defined as binary indicators that take the value 1 if a keyword appears at least once in a document, and 0 otherwise. This binary representation is adopted to stabilize estimation under sparse conditions and to focus on the presence of technological components rather than their absolute frequency. We primarily consider two representative XR target technologies, reality and extend, as response variables. In this study, two target keywords, extend and reality, are selected as representative anchors for XR technology analysis. These keywords are chosen to reflect two fundamental dimensions of XR systems, structure and integration-oriented aspects (extend) and experience and content-oriented aspects (reality). By focusing on these two dimensions, the proposed framework aims to uncover the underlying structural organization of XR technologies, rather than providing an exhaustive analysis across all possible keywords.

However, the framework is not restricted to these targets, and any keyword count can be treated as a response to investigate its conditional relationships with other sub-technologies. To reduce noise and improve stability under sparsity, we apply document frequency filtering to exclude extremely rare or overly ubiquitous keywords. Specifically, we retain keywords whose document frequency falls within a reasonable range, for example, at least 20 documents and at most 90% of documents, and we subsequently select a compact set of candidate predictors via a correlation screening step. Let $y_d$ denote the count response for a target keyword in document d. Let $z_d$ denote a vector of binary predictors indicating the presence of selected keywords, and let ${loglen}_d$ be the document length control covariate. To control for document length effects, the total keyword count for each document is included as a covariate after log transformation. This adjustment accounts for the fact that longer documents tend to contain more keywords, which may otherwise bias the estimation of keyword associations. Including document length as a control variable is essential to adjust for differences in document size, as longer patent texts tend to contain more keywords. Without this adjustment, estimated associations may be confounded by document verbosity rather than reflecting true technological relationships. We fit three likelihood-based count models. The first model is Poisson regression as follows [10,52,53].

$$y_d~Poisson\left({\mu }_d\right),log\left({\mu }_d\right)=\alpha +{\beta }_0{loglen}_d+z_d^T\beta$$

(1)

where ${\mu }_d$ is a parameter of Poisson distribution and represents a mean rate per unit interval. As follows, the second model is negative binomial regression to accommodate overdispersion [10,52,53].

$$y_d~NB\left({\mu }_d,r\right),log\left({\mu }_d\right)=\alpha +{\beta }_0{loglen}_d+z_d^T\beta$$

(2)

Our final model is zero-inflated negative binomial, which is expressed as follows [10,52,53].

$$y_d~\left\{\begin{array}{cc}0& with probability {\pi }_d\\ NB\left({\mu }_d,r\right)& with probability \left(1-{\pi }_d\right) \end{array}\right.$$

(3)

We represent the models for the two subdivided distributions as follows.

$$log\left({\mu }_d\right)=\alpha +{\beta }_0{loglen}_d+z_d^T\beta , logit\left({\pi }_d\right)={\gamma }_0+{\gamma }_1{loglen}_d$$

(4)

All models are estimated using maximum likelihood estimation (MLE) implemented in generalized linear mixed models, which supports NB and ZI components within a unified likelihood framework [47,48]. Binary indicators are used for predictor variables to reduce the influence of extreme count values and to focus on the presence or absence of technological components. This approach improves model stability and interpretability, particularly in high-dimensional sparse settings. All models were estimated using maximum likelihood estimation implemented in the R statistical environment, specifically using packages that support generalized linear models and zero-inflated modeling. Default optimization settings were used, and convergence was verified for all models. Model selection is performed using Akaike information criterion (AIC), and residual diagnostics based on dispersion and zero-inflation checks are conducted. For the conditional mean model, coefficient estimates are reported on an interpretable scale using the IRR.

$${IRR}_j=exp\left({\beta }_j\right)$$

(5)

where ${\beta }_j$ is the coefficient for predictor j. An IRR greater than 1 indicates that the presence of the corresponding keyword is associated with higher expected target counts, after adjusting for document length and other predictors. Conversely, IRR less than 1 indicates a negative conditional association. This IRR-based interpretation provides a statistically coherent measure of technology relatedness among sub-technologies. To summarize multivariate conditional relationships, we construct a directed technology relatedness network using statistically significant predictors from the selected generalized linear mixed modeling. Nodes represent sub-technologies (keywords) and targets, and a directed edge from keyword k to target t is created if the corresponding coefficient is statistically significant. Edge weight reflects association strength, for example using $\left|log\left(IRR\right)\right|$, while edge type encodes association direction (IRR ≥ 1 vs. IRR < 1). This network view complements coefficient tables by providing an intuitive structural map of which sub-technologies are conditionally linked to key XR concepts such as reality and extend. While regression quantifies conditional effects with respect to a target, technology structure in XR patents is also characterized by recurring combinations of keywords. To extract such groups, we visualize the relatedness network between XR sub-technologies. The technology relatedness network is constructed by selecting predictor variables whose coefficients are statistically significant at a predefined significance level. Directed edges are created from predictor keywords to the target keyword, and edge weights are defined based on the magnitude of the corresponding IRR values. This procedure ensures that the network reflects statistically meaningful relationships rather than simple co-occurrence patterns.

4. Experiments and Results

The objective of this analysis is to uncover the static structural relationships among XR technology keywords, rather than to model their temporal evolution. Therefore, the results should be interpreted as representing the overall organization of XR technologies across the dataset. We collected the patents related to XR technology from the KIPRIS and USPTO [50,51] for XR technology analysis. To analyze the DKM of XR technology, we used the R project and its packages [16,17,47,48,54,55]. All figures were generated with high resolution and carefully designed to ensure readability and clear presentation of the results. This experiment was conducted on the DKM constructed from XR patents. Each row in the DKM represents a patent document, each column represents a keyword representing a sub-technology, and each element represents the frequency of occurrence of that keyword. Due to the general characteristic of patent text data, the DKM is highly sparse and suffers from zero inflation. Preprocessing was performed to remove all zero rows—those with all column values set to zero. In cases where identical keywords were duplicated in columns, suffixes were removed. Then, columns with identical base names were combined to normalize the DKM. This prevented selection errors in the regression model and prevented the problem of identical sub-technology being estimated as fragmented. Additionally, to control for the document length effect, the total occurrence count $L_d=\sum _k{m}_{dk}$ was calculated for each document, and ${loglen}_d=log\left(L_d+1\right)$ was included as a covariate. This was intended to reduce the confounding effect of increased keyword counts in longer documents, thereby increasing the stability of the estimated conditional association between keywords and technologies. While the two keywords reality and extend were the primary targets of this study, the proposed framework can target any keyword, enabling a multi-faceted exploration of the structure of XR sub-technologies. Each element targeted by DKM is count data, which suffers from overdispersion as the variance is greater than the mean, and zero-inflated data. Therefore, to analyze such zero-inflated data, this study fitted and compared three likelihood-based models, Poisson regression (PR), negative binomial regression (NBR), and zero-inflated negative binomial regression (ZINBR). In this paper, model comparisons were performed based on AIC. A lower AIC indicates a better balance between data fit and complexity. In general, Poisson is disadvantaged in DKM due to its inability to account for overdispersion, while NB often significantly improves AIC by absorbing overdispersion. Furthermore, in cases where there is a structural excess of zeros, ZINB can provide additional improvements over NB. The AIC values for the three models considered in this paper are shown in

Table 1. AIC values of three models.

Target	Model
	PR	NB	ZINBR
Extend	6267.18	6257.34	6261.09
Reality	5582.91	5389.11	5393.11

below.

Table 1 reports the AIC values obtained from three count models for the two target keywords, reality and extend. Since a smaller AIC indicates a better trade-off between goodness of fit and model complexity, the results suggest that NBR provides the best fit for the reality target (AIC = 5389.11), substantially improving over PR (AIC = 5582.91). This reduction implies that the reality counts exhibit meaningful overdispersion, which is not adequately captured by the Poisson assumption (mean–variance equality), whereas the negative binomial likelihood accommodates extra-Poisson variability more effectively. For the extend target, NBR again yields the lowest AIC (AIC = 6257.34), although the difference among models is relatively modest compared to the reality case. Although the ZINB model is designed to account for excess zeros, the results indicate that it does not provide sufficient improvement over the NB model in terms of AIC to justify the additional model complexity in this dataset. This result suggests that the observed overdispersion in the data can be effectively captured by the NB model, and that the contribution of an explicit zero-inflation component may be limited under the current data conditions.

Based on these results, we adopt NBR as the primary model for subsequent inference on technology relatedness via IRR-based interpretation and downstream analyses. That is, Table 1 presents the AIC values for the three models. Since a lower AIC indicates a better balance between model fit and complexity, the results suggest that the negative binomial model provides the most appropriate fit for the data. This finding indicates that overdispersion is a key characteristic of the XR patent keyword counts and should be explicitly accounted for in the analysis. It is important to note that the objective of model selection in this study is not to optimize predictive performance, but to identify an appropriate statistical model that captures the underlying data characteristics, such as overdispersion and zero inflation. Therefore, likelihood-based criteria such as AIC are used to ensure a statistically coherent model specification.

Table 2. Parameter estimate of predictor each keyword for target keyword: extend.

Keyword	Estimate	Standard Error	Lower of CI	Upper of CI
loglen	0.5489	0.0193	0.5123	0.5881
wall	1.6208	0.1368	1.3736	1.9124
edg	1.4900	0.1255	1.2633	1.7575
electr	1.2806	0.1108	1.0808	1.5172
layer	1.3599	0.1193	1.1451	1.6151
contact	1.2339	0.1065	1.0419	1.4613
arrang	1.3449	0.0928	1.1747	1.5397
connect	1.4119	0.0882	1.2492	1.5958
surfac	1.2916	0.0747	1.1532	1.4466
structur	1.2380	0.0894	1.0747	1.4263
rotat	1.1674	0.0954	0.9947	1.3701
assembl	1.3284	0.1082	1.1324	1.5583
face	1.3685	0.1261	1.1423	1.6395
region	1.2015	0.1058	1.0110	1.4278
space	1.1367	0.0845	0.9826	1.3149
move	1.0227	0.0931	0.8556	1.2226
compon	1.2030	0.1091	1.0071	1.4372
signal	1.2387	0.1092	1.0421	1.4724
communic	1.1011	0.0955	0.9290	1.3050
posit	1.0896	0.0636	0.9719	1.2217
optic	1.0634	0.0940	0.8943	1.2645

summarizes the parameter estimates for the selected predictor keywords in the count regression model with extend as the response.

The reported estimate values are interpreted as IRRs, because the coefficients were exponentiated, $IRR=exp\left(\beta \right)$. Each predictor keyword is coded as a binary presence indicator, that is, 1 if the keyword appears at least once in a patent document and 0 otherwise. Hence, an IRR represents the multiplicative change in the expected count of extend when the corresponding keyword is present, holding all other predictors constant. A key control variable in the model is $loglen$, defined as $log\left(L_d+1\right)$, where $L_d$ is the total keyword count in document d. This covariate adjusts for document length effects, because longer texts tend to generate higher counts across many keywords, which may confound inference on keyword-specific associations. In Table 2, the IRR for $loglen$ is 0.5489. Because loglen is included primarily as a normalization control, its coefficient should not be over-interpreted as a technology effect. Rather, its role is to ensure that the IRRs of keywords reflect conditional relationships beyond the trivial effect of document length and overall verbosity. Among the predictors, wall (IRR = 1.6208) and edg (IRR = 1.4900) show the strongest positive associations with extend. Additional large IRRs are observed for connect (1.4119), layer (1.3599), arrang (1.3449), assembl (1.3284), surfac (1.2916), and electr (1.2806). Collectively, these predictors form a coherent theme that emphasizes geometric or structural representation and spatial organization, for example, wall, edge, surface, and layer, together with system integration and connectivity such as connect, electr, assembly, and arrangement. This pattern suggests that extend in XR-related patents is not merely an abstract extension concept, but is closely tied to an implementation layer involving spatial structure, surfaces, and their integration within connected systems. Several additional predictors, such as contact (1.2339), structur (1.2380), face (1.3685), and region (1.2015), indicate that extend frequently co-occurs with terms describing structural composition, contact-based interaction, and region or geometry definitions. In particular, the positive association with face may reflect either face-related sensing/interaction or the use of face in a geometric sense, for example, planar faces of surfaces. This highlights the importance of domain-aware interpretation and, where necessary, checking representative patents to disambiguate semantics. Finally, variables with IRRs close to 1 (or with confidence intervals that approach 1) indicate weaker independent contributions under the multivariable model. Overall, Table 2 provides evidence that the extend keyword is strongly associated with a sub-technology axis emphasizing spatial or structural representation and connectivity-driven integration, which is relevant for XR technology management because it points to a plausible implementation pathway for extended systems, the coupling of geometry or surfaces with connected components and structural assembly. Next,

Table 3. Parameter estimate of each predictor keyword for target keyword: reality.

Keyword	Estimate	Standard Error	Lower of CI	Upper of CI
loglen	2.5588	0.1713	2.2441	2.9176
augment	2.7976	0.1758	2.4734	3.1644
virtual	3.1558	0.1998	2.7876	3.5726
environ	1.2327	0.0805	1.0846	1.4010
content	1.2896	0.1028	1.1031	1.5076
interact	1.0836	0.0937	0.9147	1.2837
video	1.1386	0.0995	0.9593	1.3513
display	1.0965	0.0689	0.9694	1.2402
eye	0.9568	0.1033	0.7743	1.1823
scene	1.0744	0.1154	0.8704	1.3262
user	0.9407	0.0582	0.8334	1.0620
camera	0.8995	0.0866	0.7449	1.0862
object	0.9426	0.0640	0.8251	1.0769
visual	1.0502	0.0929	0.8830	1.2490
view	1.1248	0.0804	0.9777	1.2940
render	0.7749	0.0737	0.6431	0.9337
physic	0.9452	0.0840	0.7940	1.1251
captur	1.1091	0.1004	0.9288	1.3244
imag	0.8757	0.0587	0.7679	0.9986
map	0.8797	0.1017	0.7013	1.1033
mobil	1.0118	0.1188	0.8038	1.2736

reports exponentiated parameter estimates for predictor keywords in the model with reality as the response.

As in Table 2, predictors are binary presence indicators, and IRRs describe the multiplicative change in the expected reality count when a predictor keyword is present, after controlling for $loglen$ and all other included predictors. The results reveal a markedly different structure from the extend model, highlighting a sub-technology axis that is more closely aligned with XR experience concepts and content or interaction layers. The largest IRRs are observed for virtual (IRR = 3.1558) and augment (2.7976). These effects are substantial and strongly consistent with XR semantics, patents that explicitly mention virtual or augment are much more likely to exhibit higher reality counts. In practical terms, these findings indicate that reality is centrally embedded within the AR/VR conceptual core of XR patents, and that the association remains strong even under multivariable adjustment, suggesting a robust conditional relationship rather than a simple cooccurrence artifact. Beyond the AR or VR core, the positive IRRs for content (1.2896) and environ (1.2327) indicate that reality is frequently coupled with content generation/delivery and environment representation. This aligns with the view that reality in XR patents is expressed not only as an abstract label but in conjunction with the mechanisms that make XR experiences realistic: content pipelines and environmental context modeling. Additional positive associations for view (1.1248), video (1.1386), captur (1.1091), and display (1.0965) further suggest that reality is linked to an end-to-end experience stack, spanning acquisition/capture, media representation, viewpoint specification, and display output. In contrast, certain predictors exhibit IRRs below 1, most notably render (0.7749) and imag (0.8757). These negative conditional associations should not be interpreted as “rendering” or “image processing” being unimportant for XR. Rather, they likely reflect feature competition and wording substitution within patent texts under multivariable adjustment. For example, some patents may emphasize rendering-related terminology instead of using the term reality, leading to a negative conditional association once other AR/VR and content variables are included. Similarly, the relatively weak or slightly negative association for imag may indicate that reality is expressed more through concept-level XR descriptors (virtual/augment/content/environment) than through generic image-processing language, at least within the selected predictor set. Several predictors such as user (0.9407), camera (0.8995), and object (0.9426) are near or below 1, suggesting that reality in this regression specification is less directly tied to sensor/device terms and more strongly tied to the experience/content layer. Nevertheless, these terms may still play important roles in higher-order structures (e.g., bundle analysis or co-occurrence networks), and their influence can also emerge through interactions or different target choices. From a technology management perspective, Table 3 supports the conclusion that reality is anchored in the XR landscape by the AR or VR conceptual core (augment, virtual) and is operationalized through a content-and-environment pipeline (content, environ, video, captur, view, display). This suggests that future XR innovation and patent strategies centered on “reality” may be most effectively pursued through integrated developments that connect content creation, environmental context modeling, and media capture/display pipelines, rather than focusing narrowly on isolated device or low-level imaging terminology.

We visualize the conditional effects reported in Table 2 and Table 3 using forest plots of the estimated IRRs in Figure 2 and Figure 3. In both figures, each dot represents the point estimate of the IRR for a predictor keyword coded as a binary presence indicator, and the horizontal line denotes its 95% confidence interval. The vertical dashed line at IRR = 1 corresponds to no effect on the expected target keyword count. Accordingly, predictors whose confidence intervals lie entirely to the right of 1 indicate a statistically meaningful positive conditional association with the target, whereas those lying to the left indicate a negative conditional association, after adjusting for document length $loglen$ and the remaining predictors. Figure 2 shows the conditional effect (IRR) for target keyword extend.

Figure 2 corresponds to Table 2 and highlights that extend is most strongly associated with structural and integration-related sub-technologies. Specifically, the largest positive effects are observed for wall and edg, followed by connect, layer, arrang, assembl, surfac, and electr, most of which have confidence intervals clearly above 1. This pattern is consistent with Table 2 and suggests that patents emphasizing boundary or geometry and structural elements such as wall, edge, surface, and layer as well as system-level integration such as connect, electr, assembly, and arrangement tend to exhibit substantially higher expected counts of extend. The plot therefore provides an intuitive structural view that extend is positioned along an XR implementation axis involving spatial structure representation and connectivity-driven integration. In contrast, predictors whose confidence intervals overlap 1, for example, some motion or optics-related terms, show comparatively weaker conditional contributions under the multivariable model. The $loglen$ term appears as a control covariate rather than a sub-technology keyword, and its inclusion ensures that these IRR estimates represent associations beyond trivial document-length effects. Next, we represent the conditional effects (IRR) for target keyword reality in Figure 3.

Figure 3 is related to Table 3 and shows a markedly different profile, centered on XR experience-layer concepts. The strongest positive effects are clearly associated with virtual and augment, whose IRRs are substantially greater than 1 and well separated from the null line. This finding aligns with Table 3 and indicates that the presence of AR or VR core concepts is the dominant conditional driver of reality counts. Several additional predictors such as content, environ, video, view, captur, and display, tend to lie on or to the right of IRR = 1, supporting the interpretation that reality is closely linked to an end-to-end experience/content pipeline, including environmental context and media acquisition/display components. Conversely, predictors such as render and, to a lesser extent, imag appear with IRRs below 1, suggesting a negative conditional association that may reflect wording substitution and feature competition within multivariable patent text descriptions rather than a literal technological incompatibility. Overall, Figure 3 provides a compact visualization of the distinct technology axis associated with reality, contrasting with the structural or connectivity emphasis observed for extend in Figure 2. In this context, a technology stack can be interpreted as a set of interrelated sub-technologies that are jointly implemented to deliver XR functionalities. The proposed network analysis enables the identification of such stacks by revealing statistically significant relationships among keywords. We show the technology relatedness network based on significant IRR edges using the previous results as the following figure.

Figure 4 summarizes the XR technology structure by integrating statistically significant relationships derived from the regression models. Specifically, nodes represent keywords, and directed edges indicate significant conditional associations based on IRR estimates. By focusing on statistically validated relationships rather than raw co-occurrence, the network provides a more reliable and interpretable representation of XR sub-technology structure. The construction of this network follows directly from the inference pipeline summarized in Table 2 and Table 3, and visualized in Figure 2 and Figure 3. Specifically, for each target keyword of extend and reality, we fitted likelihood-based count models and reported exponentiated coefficients as IRRs, which quantify the multiplicative change in the expected target count when a predictor keyword is present, after controlling for document length $loglen$ and other predictors. Figure 2 and Figure 3 displayed these IRRs and their confidence intervals as forest plots, enabling a clear identification of predictors whose conditional effects are consistently above or below the null line IRR = 1. Figure 4 then integrates these results by retaining only statistically significant IRR edges and representing them as directed links from predictor sub-technologies to the target concepts, with edge thickness proportional to $\left|log\left(IRR\right)\right|$ and edge type indicating the sign of association $IRR\ge 1 \mathrm{v}\mathrm{s}. IRR<1$. In doing so, Figure 4 converts coefficient level evidence into an interpretable structural map of XR sub-technologies.

A key conclusion from Figure 4 is that the XR patent landscape in this dataset exhibits a dual-axis sub-technology structure anchored by two qualitatively different target concepts. The extend-centered cluster is dominated by keywords such as wall, edg, surfac, layer, structur, and integration-related terms including connect and electr. This configuration, consistent with Table 2 and Figure 2, indicates that extend is primarily associated with an implementation and integration layer of XR, emphasizing geometric or spatial structure representation, surface or edge modeling, component arrangement/assembly, and connectivity. In contrast, the reality-centered cluster, consistent with Table 3 and Figure 3, is driven by virtual and augment as the strongest predictors, alongside experience-layer and pipeline-related terms such as content, environ, and media or view-related descriptors. This pattern positions reality within an experience and content layer, where value creation is mediated by virtualization/augmentation, content generation, and environmental context representation. Importantly, the separation of these clusters suggests that XR patents operationalize extended reality not as a single monolithic technology, but as a layered stack in which distinct families of sub-technologies play different roles.

Figure 4 also clarifies the role of $loglen$ as a control covariate rather than a sub-technology. Its placement between the two target neighborhoods reflects that document length is a global source of variation that can influence keyword counts broadly. However, by explicitly controlling for $loglen$ in the regression framework, the network edges represent conditional relationships beyond trivial verbosity effects. Consequently, the observed contrast between the extend and reality neighborhoods is unlikely to be an artifact of document length and is instead indicative of meaningful, model-supported structure in the XR technology space. The network in Figure 4 provides a statistically grounded interpretation of XR sub-technology relatedness. Relatedness here is not defined by raw cooccurrence alone; rather, it reflects conditional association under multivariable adjustment. This distinction is important for technology analysis because it reduces spurious correlations driven by long documents or common background terms. The strong extend links to structural and connectivity terms imply that patents emphasizing spatial structure and system integration constitute a coherent sub-technology family. Conversely, the strong reality links to AR or VR core descriptors and content or environment terms imply another coherent family centered on immersive experience. The resulting map can therefore serve as an interpretable taxonomy for XR sub-technologies, where keywords are not only grouped by frequency but also by statistically validated relationships to central concepts. Although explicit time information is not required for the present inference, Figure 4 suggests a practical approach to technology prediction in the sense of identifying likely future integration directions. A natural prediction is that competitive innovation will increasingly occur at the interfaces between the two axes, namely where experience/content pipelines must be tightly coupled with robust spatial structure representation and system integration. From a patent strategy perspective, this implies that future high-impact XR inventions may not lie solely within the core AR or VR concept layer or solely within the structural integration layer, but in their systematic combination. Therefore, combinations that bridge the reality cluster, for example, virtual, augment, content, and environ, with the extend cluster such as surfac, layer, edg, and connect can be interpreted as promising white-space directions for R&D and intellectual property (IP) exploration, particularly if they are underrepresented in current patents despite their conceptual complementarity. The dual-axis structure supports a portfolio strategy that differentiates between experience-layer R&D and implementation/integration-layer R&D, while actively investing in cross-layer integration capabilities. For example, firms specializing in content and immersive user experience may strengthen their competitive position by partnering with or acquiring capabilities in spatial structure modeling and connectivity integration. Conversely, firms with strengths in hardware integration and spatial structure representation may gain strategic advantage by integrating advanced content pipelines and environment modeling into their platforms. In both cases, Figure 4 provides a data-driven rationale for defining XR R&D roadmaps around bundles rather than isolated keywords, and for prioritizing cross-cluster integration as a key innovation lever. From a service and deployment standpoint, the reality neighborhood aligns naturally with XR service experiences such as immersive content delivery, virtual environments, and interaction-centric applications, whereas the extend neighborhood aligns with XR system robustness and scalability such as structured spatial mapping, device configuration, connectivity and integration. This implies that end user XR services will be most competitive when they achieve both high-quality immersive experience and reliable spatial or system integration. Consequently, XR service providers may interpret the two clusters as two complementary design requirements, first, experience fidelity and content-environment richness, and second, stable spatial representation and integrated system operation.

In addition, Figure 4 demonstrates that the proposed analysis framework, combining likelihood-based count modeling with IRR interpretation and network synthesis, yields a transparent and reproducible technology map that is robust to the sparsity and zero-inflation properties of patent DKM. Unlike purely descriptive cooccurrence graphs, the presented network is explicitly grounded in a statistical model that controls for document length and isolates conditional associations. This strengthens the evidential basis of the extracted XR technology structure and supports more credible downstream decisions in technology management, including sub-technology taxonomy design, integration forecasting, and strategic white space exploration. In summary, the integrated evidence from Table 2 and Table 3, Figure 2 and Figure 3, and the network in Figure 4 indicates that XR patents in this dataset are organized around two major sub-technology axes: an AR or VR experience-and-content axis anchored by reality, and a structure and integration axis anchored by extend. The most actionable innovation opportunities are likely to emerge from cross axis integration, providing clear guidance for future XR technology forecasting, R&D prioritization, and service-oriented system design.

From a practical perspective, these findings provide valuable insights for metaverse developers. The identified dual-axis structure suggests that innovative XR systems are likely to emerge from the integration of experience-oriented components such as virtual, augment, content, environment with structure and connectivity-oriented components related to surface, layer, edge, and connect. Therefore, the developers can use this framework to identify promising technology stacks by focusing on combinations that bridge these two axes. While predictive evaluation methods such as cross-validation and out-of-sample testing can provide additional insights, they are not the primary focus of this study. The proposed framework emphasizes interpretability and structural understanding of technology relationships rather than predictive accuracy. A more detailed assessment of zero-inflation mechanisms, including formal statistical tests such as the Vuong test and residual diagnostics, may provide additional insights into the relative suitability of NB and ZINB models. However, such analyses are beyond the scope of the present study. Incorporating temporal information may provide additional insights into how XR technologies evolve over time. However, such analysis would require dynamic modeling frameworks and is beyond the scope of the present study.

5. Discussion

It is important to note that the primary objective of this study is not to demonstrate the superiority of a specific model in terms of predictive performance, but rather to identify and interpret the structural relationships among XR technology keywords. In particular, the proposed framework aims to construct an interpretable technology relatedness network, as illustrated in Figure 4, based on statistically grounded associations. While model comparison in this study is based on likelihood-based criteria such as AIC to ensure appropriate statistical fit, we do not claim that the selected model is universally superior in predictive performance compared to alternative machine learning approaches. Instead, the emphasis is placed on interpretability and statistically coherent inference for technology structure analysis. An important direction for future research is to extend the current framework by incorporating predictive performance evaluation across different statistical and machine learning models. For example, metrics such as cross-validation error, RMSE, or out-of-sample likelihood can be used to systematically compare alternative approaches. Such extensions would complement the current interpretability-focused framework and further enhance its applicability in XR technology forecasting.

The proposed framework also provides practical value for metaverse development. By identifying statistically grounded relationships among XR sub-technologies, the analysis helps developers understand which combinations of technologies are currently emphasized in patent activity. In particular, the results suggest that the most innovative technology stacks are those that integrate immersive experience components with robust structural and connectivity elements, offering guidance for system design and R&D prioritization. The future research may extend this framework by incorporating predictive performance evaluation using cross-validation, out-of-sample testing, and metrics such as RMSE or test log-likelihood. Such extensions would complement the current interpretability-focused approach and provide a more comprehensive evaluation of alternative modeling strategies. While our current analysis focuses on two representative target keywords, extend and reality, the future research may extend this framework to multiple targets or clustered keyword groups, enabling a more comprehensive exploration of XR technology structures. Also, our future research may extend the proposed framework by incorporating temporal information, such as patent filing years, to analyze the dynamic evolution of XR technologies. In particular, panel data models or time-series approaches could be used to capture changes in technology relationships over time.

6. Conclusions

This study presented a patent-driven technology analysis framework for XR that combines likelihood-based count modeling with interpretable network synthesis to map sub-technology structure from a sparse DKM. Using keyword counts extracted from XR patent documents as proxies for XR sub-technologies, we addressed the fundamental challenge that patent DKM is highly sparse and often exhibits excess zeros and overdispersion, which can degrade the reliability of conventional text-mining pipelines when applied without an explicit count data perspective. Methodologically, we modeled target keyword counts under Poisson, Negative Binomial, and Zero-Inflated Negative Binomial specifications using generalized linear mixed modeling and selected models via RMSE, complemented by simulation-based residual diagnostics. The results consistently indicated that the Negative Binomial model provided the most appropriate balance between fit and complexity for the analyzed targets, implying that overdispersion is a dominant feature of the XR patent DKM and should be accounted for in statistical inference. We then interpreted exponentiated coefficients as IRRs to quantify conditional technology relatedness, and translated statistically significant IRR effects into a technology relatedness network that provides an intuitive, structurally grounded view of XR sub-technologies.

Empirically, the integrated evidence from coefficient tables, forest plots, and the relatedness network revealed a clear dual-axis organization of XR technologies in this dataset. The target reality was anchored in an AR or VR experience-and-content axis characterized by strong associations with core concepts such as virtual and augment, and reinforced by content and environment pipeline terms. In contrast, the target extend was embedded in a structure-and-integration axis characterized by spatial or structural representation and system integration terms such as surfac, layer, edg, and connect. This separation suggests that XR patents operationalize extended reality as a layered technology stack rather than a single monolithic domain, with distinct sub-technology families corresponding to experience layer versus implementation or integration layer innovation. From a technology management perspective, the derived structure supports actionable guidance for XR R&D and IP strategy. First, the conditional relatedness network provides a statistically grounded taxonomy for identifying core sub-technology clusters and bridge components. Second, the dual-axis structure suggests that high-impact innovation opportunities are likely to arise at cross-axis interfaces, where immersive experience and content pipelines are coupled with robust spatial structure representation and system integration. Accordingly, portfolio design and R&D planning may benefit from bundling sub-technologies into coherent development tracks, experience or content vs. structure or integration, while prioritizing integration projects that connect these tracks. Moreover, the framework naturally supports white space exploration by highlighting underrepresented but conceptually complementary combinations that warrant targeted patent landscaping and expert review.

This work has several limitations that motivate future research. The current analysis is based on keyword counts from titles and abstracts; incorporating richer text fields and stronger normalization of synonyms and multiword expressions may improve semantic fidelity. In addition, the present dataset does not explicitly encode time, which restricts direct temporal forecasting; future work may extend the framework by integrating publication years and adopting dynamic count models to quantify sub-technology evolution. Finally, while IRR-based relatedness provides interpretable conditional associations, causal interpretations should be avoided; additional validation using external metadata such as assignees, international patent classification classes, or citation networks would further strengthen managerial insights. Overall, this study demonstrates that combining statistically principled count modeling with interpretable network synthesis offers a practical and reproducible approach for XR patent technology analysis under sparsity and zero inflation. The proposed framework provides an evidence-based map of XR sub-technologies, clarifies the structural separation between experience or content and structure or integration layers, and offers concrete implications for technology forecasting, R&D prioritization, and service-oriented XR system design.

Future research can extend the proposed framework in several directions. First, incorporating temporal information such as patent filing years would enable dynamic analysis of XR technology evolution and facilitate technology forecasting using time-aware models. Second, recent advances in large language models (LLMs) offer promising opportunities for improving keyword extraction, semantic representation, and contextual understanding of patent texts, which can further enhance the quality of document–keyword matrices. Third, graph-based deep learning methods, such as graph neural networks (GNNs), can be applied to model complex relationships among XR sub-technologies beyond pairwise associations. In addition, future studies may incorporate predictive performance evaluation across statistical and machine learning models, using metrics such as cross-validation error and out-of-sample likelihood, to complement the interpretability-focused approach of this study. Finally, integrating additional patent metadata, such as assignees, classification codes, and citation networks, would enable more comprehensive and multi-dimensional XR technology analysis. Recent advances in large language models (LLMs) provide promising opportunities for improving patent text analysis. In particular, LLMs can enhance keyword extraction, semantic representation, and contextual understanding of patent documents, which may complement the count-based modeling framework proposed in this study. Integrating LLM-based approaches with statistically grounded methods represents an important direction for future research. Another future research task may incorporate formal model comparison procedures, such as the Vuong test, as well as detailed diagnostics to distinguish structural and sampling zeros, in order to further investigate the role of zero-inflation in patent-derived count data.