From Traffic to Quality: A Study on the Dual-Path Driving Effects of Streamer Traits on Consumer Trust and Identification

Abstract

This study is based on the practical context of the livestream e-commerce industry’s shift from “traffic competition” to “quality competition”. Addressing the limitations of existing research that predominantly focuses on streamers’ external traits while overlooking intrinsic qualities and frequently employs linear models that oversimplify the decision-making processes of consumer purchasing behavior (CPB), a theoretical framework grounded in the Elaboration Likelihood Model (ELM) is developed to explain how streamer traits drive consumer trust and identification through dual pathways. This study adopted a mixed-method approach combining structural equation modeling (SEM) and artificial neural networks (ANNs). By analyzing 408 valid questionnaires, it systematically investigated the driving mechanisms through which streamer traits affected consumers’ trust and identification. The study found that streamers’ integrity significantly enhanced perceived trust and perceived identification via the central route. While awareness could strengthen identification, it had no significant effect on trust building, revealing the inherent tension between “traffic” and “quality”. ANN analysis further demonstrated that the nonlinear combination of traits more effectively predicts consumer responses than traits. This study provided empirical support for the “quality transformation” of livestream e-commerce from both theoretical and methodological perspectives, offering important implications for platforms to develop a quality assessment system centered on trust and identification and to optimize the streamer cultivation mechanism.

1. Introduction

Against the backdrop of the deep integration and development of the digital economy, livestream e-commerce has become a significant driver of consumption upgrades and business model innovation. As the industry shifts from extensive growth to high-quality development, the professional competence and ethical standards of streamers have increasingly become critical factors influencing the healthy development of the industry [1,2]. In recent years, incidents involving top streamers losing credibility have occurred frequently. For example, well-known streamers such as “Crazy Xiaoyang” and “Northeast Sister Yu” have drawn public attention due to false advertising and related issues. These incidents not only directly violate consumer rights but also inflict systemic damage on the industry’s credibility [3]. Fundamentally, the low entry barriers in the industry are a significant cause of disorder, with streamers’ professional competence and compliance awareness struggling to improve concurrently. Therefore, at this critical stage of livestream e-commerce development, how to cultivate a long-term trustworthy consumer environment by enhancing streamer quality and facilitating the industry’s strategic shift from “traffic-driven” to “quality-led” has become a pivotal issue for both academia and industry [4]. This transformation is not only related to the effective protection of consumer rights but also exerts a profound influence on the evolutionary direction and quality paradigm of retail in the digital economy era.

Current research on the impact of streamer traits on consumer purchasing behavior (CPB) predominantly concentrates on explicit features such as interactivity, attractiveness, and traffic [4,5,6]. However, there is a significant lack of attention to internal qualities such as integrity. Regarding the theoretical framework, existing studies primarily rely on the Technology Acceptance Model (TAM) [7] and Stimulus–Organism–Response (SOR) [8], as well as the theory of planned behavior (TPB) [9], among other paradigms, which fail to adequately consider consumers’ information processing mechanisms. In fact, current research methods predominantly depend on structural equation modeling (SEM), which is effective in validating linear relationships among variables, but exhibits clear limitations in identifying complex nonlinear interaction effects, thereby oversimplifying consumers’ actual decision-making processes [10]. Accordingly, this study incorporates an artificial neural network (ANN) with adaptive learning capabilities to thoroughly investigate the latent nonlinear mechanisms and complex quantitative relationships between streamer traits and CPB, transcending the theoretical constraints of traditional linear models and offering more refined mechanistic insights into CPB.

To address the aforementioned research gaps and respond to the practical needs of the livestream e-commerce industry’s quality transformation, this study makes three core innovative contributions to the existing literature. First, it innovatively introduces the Elaboration Likelihood Model (ELM) into the research of livestream e-commerce and constructs a dual-path theoretical framework for streamer traits to influence consumer psychology and behavior, filling the gap that traditional theories ignore the heterogeneity of consumer information processing. Second, it systematically clarifies the theoretical rationale for mapping specific streamer traits onto central and peripheral routes based on the cognitive effort required for consumer information processing and the attribute characteristics of streamer traits: professionalism, interaction, and integrity are defined as central route core cues because they reflect streamers’ intrinsic competence and moral attributes, requiring consumers to invest in in-depth observation, rational judgment, and cognitive processing to form reliable evaluations; reputation and attractiveness are categorized as peripheral route heuristic cues because they represent streamers’ external social attributes and sensory characteristics, allowing consumers to form intuitive perceptions and emotional responses through simple and superficial information processing without in-depth deliberation. Third, it adopts a mixed-method approach combining SEM and ANN, which not only verifies the linear driving effects of single streamer traits on consumer trust, identification, and purchase behavior but also explores the nonlinear combination effects of multiple traits, making up for the deficiency of existing research that only focuses on linear relationships and ignores the complex interaction of traits. Based on this, this study analyzes 408 valid questionnaires to systematically investigate the underlying mechanisms through which streamer traits drive consumer trust and identification via dual pathways, and further compares the predictive power of linear and nonlinear effects on CPB, aiming to provide a more comprehensive theoretical explanation and empirical evidence for the industry’s quality transformation.

The findings demonstrate that streamer integrity significantly enhances consumers’ perceived trust and perceived identification through the central route. While awareness can improve the sense of identification, it does not have a significant effect on trust formation. This finding exposes the inherent tension between “traffic” and “quality”. ANN analysis further demonstrates that the nonlinear combination effect among traits predicts consumer decision-making more effectively than individual traits.

2. Literature Review

2.1. Research on Streamer Traits

Streamers, as online opinion leaders emerging through social media, serve as key nodes within the livestream social ecosystem and significantly influence CPB [11,12]. A synergistic evolutionary and mutually reinforcing relationship exists between streamer traits and the development of the livestream e-commerce industry.

In promoting the sustainable and healthy development of livestream e-commerce, the multidimensional traits possessed by streamers play a crucial role. Existing studies typically characterize streamer traits in terms of professionalism, interactivity, awareness, and attractiveness [7,13,14]. Empirical studies demonstrate that these traits have a significant positive effect on CPB. Specifically, streamers’ professionalism directly shapes consumers’ trust in the products and their purchase intentions [3,13]. Interaction ability profoundly influences the atmosphere construction within the livestream environment and the final conversion outcomes [13]. Furthermore, streamers with high awareness are more likely to obtain consumers’ initial trust and identification, thereby promoting the adoption of recommended products [15]. From the perspective of social cognition, consumers generally display more positive attitudes with high social attractiveness [7]. These traits offer robust theoretical support for the mechanisms through which streamer traits affect CPB.

However, existing literature seldom focuses on integrity. Integrity is characterized by authentic and credible self-presentation and consistency between words and actions, which can enhance emotional identification and sustained willingness to follow, thereby elevating the industry’s credibility and influence through positive word-of-mouth dissemination. Notably, existing research lacks a clear and systematic classification framework for streamer traits based on the logic of consumer information processing, and fails to distinguish the differences in cognitive processing modes required for different traits, which makes it impossible to accurately reveal the differentiated influence mechanisms of different streamer traits on consumer psychology and behavior. This study fills this gap by combining ELM to classify streamer traits into central and peripheral route cues, and clarifies the theoretical basis for classification from the perspective of information processing depth and cognitive effort input, which helps to deepen the understanding of the influence mechanism of streamer traits.

2.2. Elaboration Likelihood Model

ELM, proposed by Petty et al. in social psychology research, has been widely applied in studies on individual attitude formation, behavior mechanisms, and the impact of social communication [16]. The theory posits that changes in individual attitudes are primarily achieved through the central route and the peripheral path. The central route relies on rational deliberation, requiring individuals to invest significant cognitive effort and time to systematically process core information directly related to the task. The peripheral route, in contrast, depends on simple and heuristic external cues, making the process more time-efficient and requiring no in-depth cognitive engagement or systematic information analysis.

As a classic theory of consumer information processing and attitude change, the core value of ELM lies in its ability to distinguish two different information processing pathways and their respective influencing factors, which makes it particularly suitable for explaining consumer decision-making behavior in the livestream e-commerce context—an environment where consumers are simultaneously exposed to rational and substantive information such as streamers’ professional product explanations and emotional and superficial cues such as streamers’ personal attractiveness and social reputation. Based on the core connotation and theoretical assumptions of ELM, the classification of streamer traits into central and peripheral route cues in this study follows a clear and rigorous theoretical rationale: on the one hand, traits that reflect streamers’ intrinsic competence and moral attributes (professionalism, interaction, integrity) belong to central route cues. Professionalism requires consumers to judge the streamer’s product knowledge and professional ability through in-depth listening to product explanations; interaction needs consumers to perceive the streamer’s sincerity and attention through sustained two-way communication; integrity demands consumers to identify the streamer’s moral quality through the consistency of their words and deeds and the authenticity of product recommendations—all of which require consumers to invest a certain amount of cognitive effort for in-depth processing and judgment. On the other hand, traits that reflect streamers’ external social attributes and sensory characteristics (reputation, attractiveness) belong to peripheral route cues. Consumers can easily obtain information about a streamer’s reputation through the number of followers and popularity, and can directly perceive the streamer’s attractiveness through appearance and personal charisma—these cues can be processed heuristically with minimal cognitive effort, and consumers form intuitive emotional perceptions and judgments based on these cues.

Because ELM accurately depicts the situational factors and psychological mechanisms underlying attitude formation, the model is widely employed in research fields such as e-commerce, marketing, and information systems to explain individual behavioral decision-making [17,18]. For example, in e-commerce research, Wang et al. employed the central route and peripheral path to elucidate the evolutionary mechanisms of CPB in the context of homogeneous competition [2]. In the field of marketing, Jayawardena et al. designed a qualitative experiment based on the ELM, comparing the differentiated persuasive effects of luxury and non-luxury print advertisements on consumers’ brand attitudes. These results demonstrate that this model not only provides a theoretical foundation for the stage demarcation of attitude changes within social cognitive processes, but its conclusions have also been applied in practice to guide attitude transformation [19].

Compared with other consumer behavior theories, ELM has the following notable advantages: firstly, it reveals the internal influence mechanism through which external information affects individual cognition and behavior via different pathways. Secondly, it explains why the same influence mechanism produces differential effects on different users under the same external environment [7,20]. In the field of livestream e-commerce research, the application of ELM can make up for the deficiency of traditional theories that ignore the heterogeneity of consumer information processing, and provide a more refined and systematic theoretical framework for revealing the differentiated influence of different streamer traits on consumer trust, identification, and purchase behavior.

2.3. Research on Consumer Purchase Behavior

CPB refers to the sequence of decisions and activities through which consumers select, purchase, use, and evaluate products and services to satisfy their needs [21]. An accurate understanding of CPB is essential for enterprises to develop effective marketing strategies, as it directly influences product market performance and brand competitiveness.

In the field of CPB, SOR [8], TAM [9], and TPB [22] are classical theories that provide important frameworks for understanding consumer decision-making. However, these models still have limitations in elucidating the intrinsic mechanisms of individual information processing and attitude formation, thereby hindering a comprehensive interpretation in complex contexts [2,23]. In fact, CPB is influenced by multiple dimensions, including personal factors (such as age, gender, occupation, and income), psychological factors (such as motivation, attitude, and perceived trust), emotional factors (such as emotion and mood), and environmental factors (such as product and streamer) [5,13,24]. Notably, a significant association exists between streamer traits and CPB [3,7,13]. Streamers subtly influence consumers’ decision-making psychology and behavioral tendencies through various approaches, including professional expertise, interactive capability, emotional resonance, and trust construction.

At the methodological level, existing literature generally employs an empirical research paradigm for analysis. For instance, Faqih [25] validated the significant effects of factors such as gender, trust, and perceived risk on CPB through the construction of SEM. Nonetheless, traditional SEM exhibits certain limitations in addressing complex variable relationships, particularly in effectively capturing potential nonlinear interactions among variables. To overcome this limitation, Wang et al. (2025) introduced a hybrid modeling approach whereby, after identifying key independent variables using SEM, ANN was further incorporated to thoroughly investigate nonlinear relationships among variables and assess predictive performance [2]. Another recent study by Ji et al. (2025) empirically examined the combined influence of streamer attractiveness and product appearance on consumer purchase decisions using an eye-tracking experiment [5]. Furthermore, some scholars have begun to explore the integration of advanced computational techniques, such as deep learning, to extract deeper behavioral patterns and influence mechanisms from multi-source data [26,27].

3. Research Hypotheses and Research Model

3.1. Research Hypotheses

3.1.1. Effects of Central Route Variables on Perceived Trust and Perceived Identification

(1) The Relationship between Professionalism and Perceived Trust and Perceived Identification

Professionalism refers to the depth of knowledge, skill proficiency, and relevant experience demonstrated by the streamers during the presentation and explanation of products [28]. It not only encompasses professional cognition of the product itself but also includes credible factors such as experience and qualification certification. Streamers with higher professional expertise are more capable of effectively controlling the livestream pace and creating a trustworthy marketing context [29]. Perceived trust refers to the psychological expectation consumers form regarding the reliability and credibility of the recommended content, based on their judgments of the streamers’ ability, integrity, and benevolence [3]. Products recommended by streamers with stronger professional expertise tend to possess greater persuasive power, which helps rapidly alleviate consumers’ perceived risk and decision-making doubts, facilitating smooth transaction completion [29,30]. Professionalism, as a high-cost and difficult-to-imitate credible signal, can effectively reduce information asymmetry, thereby enhancing the evaluation of the credibility of the streamers’ recommended content [31,32]. Related research also demonstrates that professionalism is a key driver in consumers’ formation of initial trust [30,31].

Perceived identification refers to the psychological connection and sense of belonging that individuals establish with others in terms of values, emotion, or identity [33]. Within livestream contexts, consumers are more inclined to develop value identification and a sense of identity with streamers who exhibit a high level of professional expertise, as professional competence signifies reliability and serves as a benchmark for social value. Simultaneously, based on the central route mechanism of ELM, professional expertise, as a high-quality informational cue, can elicit consumers’ deep cognitive processing, thereby enhancing their identification with the streamers’ professional image [3]. Magni & Rossi (2017) also indicated that professional qualifications and knowledge reserves were important antecedents of consumers’ value identification [34]. Therefore, the following inferences can be drawn:

H1a. 

Professionalism positively affects perceived trust.

H1b. 

Professionalism positively affects perceived identification.

(2) The Relationship between Interaction and Perceived Trust and Perceived Identification

Interaction is reflected in highly engaged social behaviors, such as real-time responses, personalized interactions, and two-way communication [35]. According to social exchange theory, streamers convey attention and respect to consumers through active interaction. This engagement triggers consumers’ reciprocity tendencies, thereby enhancing their assessment of the streamers’ credibility [31]. At the same time, sustained and consistent interactive behavior, as a visible commitment signal, can effectively reduce information asymmetry and communicate the streamers’ reliability [29]. Within the livestream context, interaction also fosters emotional connections by enhancing social presence, and according to the dual-process model of trust construction, these emotional connections, combined with cognitive evaluations, jointly facilitate the development of trust [36].

Streamers convey signals of attention and respect through active interaction, stimulating consumers’ reciprocity tendencies and emotional resonance, thereby strengthening their sense of identification with the streamers [37]. Streamers elicit consumers’ positive emotional responses by fostering a pleasant communication atmosphere and social affinity, thereby enhancing consumers’ acceptance of and identification with the streamers’ identity and values. Hence, it is reasonable to infer that the higher the streamers’ level of interaction, the stronger the consumers’ perceived identification with them. This theoretical inference is consistent with existing research findings. For instance, relevant studies demonstrate that frequent, high-quality interaction is a key predictor of identification with the streamers [38]. Accordingly, this study proposes the following hypotheses:

H2a. 

Interaction positively affects perceived trust.

H2b. 

Interaction positively affects perceived identification.

(3) The Relationship between Integrity and Perceived Trust and Perceived Identification

Integrity represents a concentrated manifestation of the streamers’ principle of honest product recommendation [39]. According to trust construction theory, integrity, as the core dimension of moral credibility, can convey consistent behavioral expectations to consumers and significantly reduce their perceived risk. Moreover, integrity, as a moral signal that is difficult to replicate, effectively mitigates information asymmetry and enhances consumers’ assessment of the streamers’ sincerity in recommendation motives. Within the highly uncertain context of livestream e-commerce, the streamers’ rejection of false advertising and truthful disclosure of product information not only aligns with consumers’ moral expectations but also facilitates trust formation by fostering a perception of emotional security [39]. Morgan and Hunt (1994) demonstrated in their seminal study that integrity is a key antecedent of trust construction [40]. Recent studies also confirm that streamers’ moral behavior decisively influences the formation of consumer trust [39].

Integrity is reflected in adherence to ethical standards during livestreaming, maintaining consistency between words and actions, and exhibiting genuine honesty [41]. Relevant research confirms that moral traits are critical predictors of consumers’ emotional identification [42]. Specifically, consumers tend to identify with entities that align with their own values and moral standards. The integrity exhibited by streamers can elicit consumers’ value resonance, thereby establishing a psychological connection based on moral identification. Streamers’ moral consistency across different contexts enhances the predictability of their image, which in turn promotes the formation of a sense of identification [43]. In the livestream e-commerce environment characterized by information asymmetry, the streamers’ integrity not only satisfies consumers’ fundamental need for transaction security but also stimulates deep identity through value congruence. Therefore, the following propositions are proposed:

H3a. 

Integrity positively affects perceived trust.

H3b. 

Integrity positively affects perceived identification.

3.1.2. The Impact of Peripheral Path Variables on Perceived Trust and Perceived Identification

(1) The Impact of Reputation on Perceived Trust and Perceived Identification

Reputation as a salient indicator of social influence and public recognition significantly promotes the establishment of consumers’ perceived trust [38]. In the highly information-asymmetric environment of livestream e-commerce, reputation becomes a crucial heuristic cue for consumers to quickly assess credibility [36]. A higher reputation, as an external signal that is difficult to fabricate, can effectively reduce consumers’ information search costs and perceived risk, thereby enhancing their initial trust in the recommended content [15]. According to institutional trust theory, the social recognition and public endorsement carried by reputation constitute a non-institutional trust assurance mechanism, making consumers more inclined to trust the professional judgment and integrity of prominent streamers. From the perspective of cognitive psychology, the “exposure effect” and “halo effect” triggered by reputation induce consumers’ cognitive shortcuts, leading them to transfer their positive feelings of familiarity into trust judgments [28].

Research indicates that the reputation of public figures significantly enhances audiences’ value identification through internalization processes [44]. According to social identity theory, consumers tend to enhance their self-image and social identity by associating with high-reputation individuals. This psychological mechanism, based on “status projection”, prompts consumers to convert their recognition of the streamer’s social reputation into internal identity [15]. In the socialized context of livestream e-commerce, focusing on a prominent streamer constitutes a declaration of social identity, which further reinforces the sense of identification with the streamer group through self-categorization mechanisms. Based on this, we propose the following hypotheses:

H4a. 

Reputation positively affects perceived trust.

H4b. 

Reputation positively affects perceived identification.

(2) The Effect of Attractiveness on Perceived Trust and Perceived Identification

Attractiveness, as a multidimensional trait encompassing external appearance, charisma, and social affinity, uniquely promotes the establishment of consumers’ perceived trust [5,45]. According to the affect transfer model, the emotional pleasure elicited by attractiveness transfers to trust evaluations of the streamer through a conditioned reflex mechanism [35]. From the perspective of evolutionary psychology, attractiveness functions as a potential cue for health and quality, inherently embodying instinctive judgment of the reliability of cooperative partners since ancient times. Effective image management inherently conveys the streamers’ professional dedication and self-imposed standards. Such high-cost signals effectively reduce consumers’ perceptions of uncertainty. In the highly visualized context of livestream e-commerce, streamer attractiveness not only enhances content enjoyment but also strengthens perceived trustworthiness through the dimension of social attractiveness [29,32].

According to self-extension theory, consumers tend to incorporate the values and lifestyles represented by attractive streamers into their self-identity construction, thereby enriching and expanding their self-concept through this psychological projection. Schouten et al. (2020) confirmed this perspective, arguing that influencer attractiveness significantly promotes consumer identity through a self-enhancement mechanism [46]. From the perspective of emotional resonance, the emotional pleasure elicited by attractiveness enhances consumers’ willingness to engage, thereby facilitating their internalized identity with the streamers [47]. In the dual contexts of visualization and socialization within livestream e-commerce, streamer attractiveness not only improves the viewing experience but also functions as a crucial symbolic resource for consumers to construct social identity through dual pathways of aesthetic resonance and value resonance.

H5a. 

Attractiveness positively influences perceived trust.

H5b. 

Attractiveness positively influences perceived identification.

3.1.3. The Influence of Perceived Trust and Perceived Identification on CPB

According to the trust–action theory, when consumers establish stable trust in a streamer, a behavioral intention based on reliable expectations naturally forms, significantly reducing risk perception and uncertainty during the decision-making process [28]. From a rational choice perspective, trust offers consumers an efficient cognitive shortcut by reducing transaction costs and information search costs. Simultaneously, based on the emotion–behavior linkage mechanism, the emotional security triggered by trust transforms into positive purchase intention and ultimately manifests as actual purchase behavior. In the highly uncertain environment of livestream e-commerce, perceived trust not only directly facilitates purchase decisions but also optimizes the entire consumer purchase process by enhancing brand favorability and reducing decision conflicts [35,47].

Perceived identification, as a psychological connection and value resonance established between consumers and streamers, exerts a profound driving effect on purchase behavior. According to social identity theory, when consumers identify with streamers, they internalize the streamers’ values and lifestyles as part of their self-concept. This identification transformation motivates consumers to maintain and reinforce this identity consistency [48]. From the emotion–behavior linkage perspective, the emotional sense of belonging triggered by identification can transform into supportive consumption motivation, significantly enhancing purchase intention and reducing price sensitivity. The study by Casidy et al. (2019) confirmed this view, finding a significant positive correlation between identification and willingness to pay [49]. Therefore, we propose the following hypotheses:

H6. 

Perceived trust positively influences CPB.

H7. 

Perceived identification positively influences CPB.

3.1.4. The Mediating Role of Perceived Trust and Perceived Identification

Perceived trust plays a critical mediating role between streamer traits and CPB, a mechanism that can be systematically explained from multiple theoretical perspectives [30]. According to the streamer trust framework, consumers utilize streamer traits to establish trust, which subsequently influences purchase behavior [50]. Specifically, traits such as the streamers’ professionalism, interaction, and integrity serve as credibility cues that are first cognitively processed by consumers via the central route, forming trust judgments regarding the streamers’ professional competence and moral qualities, whereas traits like reputation and attractiveness function as peripheral cues, eliciting consumers’ intuitive trust through the emotional pathway. When these diverse traits jointly influence consumers, perceived trust, as the core psychological mechanism, transforms trait cognition into behavioral intention, ultimately facilitating purchase decisions. This theoretical proposition is supported by multiple empirical studies. Commitment–Trust Theory provided a foundational theoretical basis for the mediating role of trust [40]. Zhu et al. (2021) demonstrated that trust served as a key mediator in the influence of streamer traits on consumer decision-making [35]. Chen et al. (2023) confirmed the mediating effect of perceived trust between professionalism, interactivity, and CPB [32]. Hence, the following hypotheses are proposed:

H8a. 

Perceived trust mediates the relationship between professionalism and CPB.

H8b. 

Perceived trust mediates the relationship between interactivity and CPB.

H8c. 

Perceived trust mediates the relationship between integrity and CPB.

H8d. 

Perceived trust mediates the relationship between reputation and CPB.

H8e. 

Perceived trust mediates the relationship between attractiveness and CPB.

From the perspective of identity consistency motivation, consumers actively seek streamer traits that align with their self-concept and internalize these traits as part of their self-identity. Escalas and Bettman (2019) further confirmed that consumers tend to establish identification relationships through celebrities and self-expression [51].

When consumers identify with streamers, they sustain and reinforce this identity consistency through purchase behavior, achieving a full transition from psychological identification to behavioral expression. Specifically, core traits of streamers, such as professionalism, interactivity, and integrity, are cognitively processed by consumers through the central route, forming value identification based on competence and morality, whereas peripheral traits, such as attractiveness and reputation, elicit consumers’ emotional resonance and identity aspirations via the emotional pathway. In this process, perceived identification, as a key psychological mechanism, transforms static trait cognition into dynamic behavioral motivation, establishing a complete effect path of “trait cognition—psychological identification—purchase behavior” [52]. Therefore, we propose the following hypotheses:

H9a. 

Perceived identification mediates the relationship between professionalism and CPB.

H9b. 

Perceived identification mediates the relationship between interactivity and CPB.

H9c. 

Perceived identification mediates the relationship between integrity and CPB.

H9d. 

Perceived identification mediates the relationship between reputation and CPB.

H9e. 

Perceived identification mediates the relationship between attractiveness and CPB.

3.2. Research Model Construction

This study, grounded in ELM, develops a model of consumer purchase behavior in the context of livestream e-commerce to analyze consumers’ dual information-processing mechanisms in Figure 1. Specifically, professionalism, interaction, and integrity serve as core cues of the central route, enhancing argument quality to promote consumers’ deep cognitive processing and rational evaluation. Reputation and attractiveness serve as heuristic cues within the peripheral path, primarily eliciting consumers’ emotional responses and affective decision-making. To systematically reveal the intrinsic transmission mechanism from streamer traits to CPB, this study introduces perceived trust and perceived identification as key mediating variables, aiming to elucidate how the two paths differentially drive the psychological mechanisms, thereby influencing decision outcomes. This model not only integrates the dual-path framework of ELM but also deepens the theoretical understanding of the mechanism underlying streamer influence by delineating the specific effects of different traits on the mediating variables.

4. Research Methods

4.1. Two-Stage Research Method Based on SEM-ANN

To systematically analyze the complex interaction mechanisms among variables within the model, this study employs a two-stage mixed-method approach combining SEM and ANN, as shown in Figure 2. This approach aims to leverage the advantages of both methods synergistically. In the first stage, SEM is used to test linear path relationships among variables in the theoretical model, identify key predictive variables that significantly influence CPB, and validate theoretical hypotheses. However, SEM is limited in capturing nonlinear relationships that may exist within complex decision-making behaviors [2]. Therefore, in the second phase, this study uses the significant variables identified by SEM as inputs and incorporates an ANN for further analysis. ANN, owing to its robust pattern recognition and function approximation abilities, can effectively process the complex linear and nonlinear interactions among variables [53]. Accordingly, it predicts the probability of CPB and ranks the relative importance of critical influencing factors. This hybrid methodology integrates the strengths of SEM for theoretical validation and ANN for predictive modeling, thereby not only reinforcing the model’s theoretical explanatory power but also significantly improving prediction robustness and accuracy [54,55].

4.2. Questionnaire Design

The questionnaire mainly comprises two parts: (1) Demographic information, including respondents’ gender, age, education level, income, and years of livestream viewing. (2) Measurement items of the core variables, all adapted from validated scales in existing literature. Variables are measured using mature items such as professionalism [39], interactivity [36], integrity [39], attractiveness [45], integrity [36], perceived trust [56,57], perceived identification [58], and CPB [59]. Measurement was conducted using a seven-point Likert scale, where 1 indicates “strongly disagree,” and 7 indicates “strongly agree,” reflecting increasing levels of attitudinal agreement.

A pilot survey was conducted to pre-test the reliability and validity of the questionnaire before formal data collection. Strictly following the principles of voluntary participation and anonymity, the pilot survey was conducted face-to-face from 1 June to 15 June 2025, and the respondents were 30 e-commerce major students from a university in the eastern region of China, all of whom had at least 6 months of livestream viewing and actual shopping experience to ensure their ability to evaluate the questionnaire items. The reliability and validity analysis of the pilot survey data demonstrates that the scale overall exhibits satisfactory reliability and validity, with all Cronbach’s α coefficients greater than 0.7 and factor loadings of all items greater than 0.6, which conforms to the basic academic measurement standards. Based on the pilot survey results, a small number of questionnaire items were revised for wording clarity to form the final formal questionnaire.

4.3. Questionnaire Collection

4.3.1. Recruitment and Inclusion/Exclusion Criteria

The formal data collection was conducted from 20 June to 30 July 2025, and the survey targeted active users with actual livestream e-commerce consumption experience in China. The specific recruitment criteria were: (1) having the ability to independently complete the questionnaire (aged 18 and above, with basic reading and understanding ability); (2) having at least 3 months of livestream viewing experience on mainstream Chinese livestream e-commerce platforms. The exclusion criteria were: (1) respondents without actual livestream shopping experience; (2) respondents with cognitive or language disorders that prevent them from completing the questionnaire; (3) respondents who participate in the survey for the purpose of obtaining rewards and have random answering behavior.

4.3.2. Sampling and Data Collection Methods

A combination of stratified random sampling and convenience sampling was adopted for this study, and online and offline mixed data collection methods were employed to improve the representativeness of the sample. A combination of online and offline data collection methods was employed: offline data were collected through one-on-one interview-based questionnaire administration, while online distribution was conducted via targeted dissemination through social media channels such as WeChat and QQ. All participants were required to rate the scales based on their actual livestream shopping experiences and subjective evaluative behaviors to ensure the authenticity of the data.

4.3.3. Questionnaire Screening and Invalid Response Filtering

A total of 415 questionnaires were distributed in this study. After the data collection, strict data cleaning and invalid questionnaire screening were conducted, and the specific screening criteria and thresholds for invalid responses were as follows: (1) questionnaires with a completion time of less than 90 s (the average time required to complete the questionnaire in the pilot survey was 150 s), which were judged as hasty answering; (2) questionnaires with obvious patterned scoring (e.g., all items rated 1, 7, or regular alternating scores such as 1-2-3-4), which were judged as random answering; (3) questionnaires with missing values of more than 10% of the total items, which could not be supplemented and corrected. After screening, the effective response rate was 98.3%. This mixed sampling and data collection strategy helps to reduce sampling bias and response bias, thereby enhancing the sample representativeness and data authenticity.

5. Data Analysis and Results

5.1. Descriptive Analysis

First, a statistical analysis of the demographic characteristics of valid samples was conducted, with the results presented in

Table 1. Demographic characteristics of the sample.

Name	Option	Frequency	Proportion
Gender	Male	162	39.71%
	Female	246	60.29%
Age	Under 18	10	2.45%
	18–25 years	107	26.23%
	26–30 years	146	35.78%
	31–40 years	107	26.23%
	41–50 years	26	6.37%
	51 years and older	12	2.94%
Income	0–2000 yuan	73	17.89%
	2000–4000 yuan	203	49.75%
	4000–8000 yuan	104	25.49%
	8000–10,000 yuan	28	6.86%
	10,000–20,000 yuan	0	0%
	Above 20,000 yuan	0	0%
Educational background	High school or below	26	6.37%
	Junior college	71	17.40%
	Bachelor’s degree	269	65.93%
	Master’s degree or above	42	10.29%
Years of viewing	Less than 1 year	30	7.35%
	1–3 years	193	47.30%
	3–5 years	126	30.88%
	More than 5 years	59	14.46%

. Regarding gender distribution, 162 male respondents accounted for 39.71% of the total sample, while 246 female respondents accounted for 60.29%, indicating a relatively higher willingness to participate among females. By age, most respondents were concentrated in the 18–25 and 26–30 age groups, representing the core demographic of online consumers. Regarding income, respondents with a monthly income between 2000 and 4000 yuan accounted for 50% of the total, indicating that the middle- to low-income group constituted the primary source of the sample for this study. The educational level is predominantly undergraduate, which generally corresponds with the age and educational structure of the current mainstream e-commerce consumer demographic. Regarding the livestream viewing experience, most respondents have 1–3 or 3–5 years of experience, indicating that the sample as a whole has substantial experience with livestream e-commerce.

5.2. Common Method Bias Analysis

To control the impact of common method bias (CMB), measures were adopted at both the procedural design and statistical testing stages. Regarding procedural control, multiple effective measures were adopted in the questionnaire design and data collection stage to reduce CMB: (1) The questionnaire design employs a multidimensional scale structure, and the measurement items of independent variables, mediating variables, and dependent variables are arranged in a scattered manner to avoid respondents’ associative answering. (2) During the questionnaire administration, respondents are explicitly informed that there are no right or wrong answers to the questions, and their answers will only be used for academic research to reduce social desirability bias. (3) The anonymity of respondents is strictly maintained, and no personal identifying information (e.g., name, ID number, phone number) is collected to eliminate respondents’ concerns about answer disclosure. (4) A mixed data collection mode combining online and offline methods is adopted, and stratified random sampling is used to expand the sample source, reducing the systematic errors caused by a single data collection method. (5) Some reverse-coded items are added to the scale to prevent respondents from forming answering inertia and improve the accuracy of the answers.

At the stage of statistical testing, the Harman single-factor analysis method was initially applied to assess CMB. Results in

Table 2. Total variance explained.

Components	Initial Eigenvalues			Sum of Squared Loadings Extracted
	Total	Variance Percentage	Cumulative	Total	Variance Percentage	Cumulative
1	11.770	37.969	37.969	11.770	37.969	37.969
2	1.891	6.101	44.070	3.098	11.473	34.701
3	1.665	5.373	49.443	2.855	10.574	45.275
4	1.460	4.709	54.151	2.029	7.514	52.789
5	1.309	4.224	58.375	1.499	5.552	58.341
6	1.267	4.088	62.463	1.295	4.795	63.136
7	1.088	3.511	65.974	1.088	3.511	65.974
8	0.702	2.263	68.237
9	0.669	2.158	70.395
10	0.632	2.038	72.433

reveal that the first factor accounts for 37.969% of the variance, falling short of the 40% threshold for total variance explained, indicating that common method bias is not a serious concern [2,53]. However, it should be acknowledged that the Harman single-factor test has inherent limitations: it is a relatively simple, preliminary CMB detection method that can only assess potential common method bias at a general level and cannot fully eliminate the influence of latent CMB. In addition, the variance explained by the first factor (37.969%) is close to the 40% threshold, suggesting the possibility of a small amount of potential CMB in the study. Therefore, the research conclusions should be interpreted with caution, and subsequent research can employ multiple data-collection methods (e.g., combining questionnaire surveys with objective transaction data) to mitigate the influence of common method bias further.

5.3. Reliability and Validity Assessment

Reliability reflects the consistency and stability of the measurement scale. This study employs Cronbach’s alpha (α) and Composite Reliability (C.R.) as indicators to evaluate reliability [60]. Partial Least Squares (PLS) analysis was performed using SmartPLS 3.0 software to obtain the outer loadings and reliability indices.

The results in

Table 3. Variables’ factor loading, VIF, α, C.R., and AVE.

Variables	Items	Factor Loading	$\mathbf{V}\mathbf{I}\mathbf{F}$	$\mathsf{\alpha }$	$\mathbf{C}.\mathbf{R}.$	$\mathbf{A}\mathbf{V}\mathbf{E}$
Professionalism	P1	0.822	1.912	0.841	0.894	0.677
	P2	0.838	2.016
	P3	0.813	1.828
	P4	0.818	1.892
Interaction	I1	0.799	1.616	0.806	0.873	0.632
	I2	0.802	1.684
	I3	0.797	1.699
	I4	0.781	1.570
Integrity	IN1	0.808	1.782	0.826	0.884	0.657
	IN2	0.820	1.857
	IN3	0.791	1.647
Attractiveness	IN4	0.823	1.760	0.808	0.874	0.634
	A2	0.823	1.798
	A3	0.809	1.716
	A4	0.743	1.489
Reputation	R1	0.830	2.012	0.847	0.897	0.685
	R2	0.822	1.803
	R3	0.839	2.115
Perceived identification	PI1	0.779	1.628	0.836	0.890	0.670
	PI2	0.836	1.967
	PI3	0.834	1.875
	PI4	0.824	1.921
Perceived trust	PT1	0.855	2.113	0.856	0.902	0.698
	PT2	0.820	1.844
	PT3	0.835	1.928
	PT4	0.833	1.979
CPB	CPB1	0.828	1.618	0.802	0.884	0.717
	CPB2	0.844	1.737
	CPB3	0.868	1.903

show that all measurement items have factor loadings (FLs) greater than 0.7, both alpha coefficients and C.R. values exceed the 0.7 threshold, and the Average Variance Extracted (AVE) values are all above 0.5, indicating that the scales demonstrate satisfactory reliability [60]. Furthermore, to eliminate the interference of multicollinearity on model estimation, the variance inflation factor (VIF) of each variable was further examined. Assessment of variance inflation factors shows that all constructs have VIF values below the 3.0 threshold. No severe multicollinearity was detected, thereby supporting the validity of the proposed model.

Discriminant validity refers to the degree to which a variable is distinct from other variables. This study utilizes the Heterotrait–Monotrait ratio (HTMT) for assessment. The smaller the HTMT values, the higher the discriminant validity between variables [61]. As shown by the values on and above the diagonal in

Table 4. HTMT values.

Variables	Attractiveness	Reputation	Interaction	Integrity	Professionalism	Perceived Identification	Perceived Trust	CPB
Attractiveness	0.796	0.373	0.489	0.513	0.463	0.567	0.526	0.576
Reputation		0.827	0.396	0.483	0.428	0.505	0.433	0.429
Interaction			0.795	0.442	0.437	0.545	0.484	0.542
Integrity				0.810	0.392	0.532	0.529	0.559
Professionalism					0.823	0.502	0.487	0.551
Perceived identification						0.819	0.682	0.550
Perceived trust							0.847	0.701
CPB								0.836

, all variables’ HTMT values are below the evaluation criterion of 0.9, indicating strong discriminant validity among these eight variables.

5.4. Hypothesis Testing

(1) Main Effect Testing

This study employs partial least squares structural equation modeling (PLS-SEM), with path coefficient estimation and hypothesis testing performed using SmartPLS 3.0 software [2]. Bootstrap resampling was set to 5000 iterations to examine the significance level of each path.

The test results are summarized in

Table 5. Main effect path analysis results.

Hypothesis	Paths	Path Coefficient (β)	Sample Mean	Standard Deviation	$\mathbf{T}$	$\mathbf{P}$
H1a	Professionalism → Perceived trust	0.247	0.245	0.046	5.410	0.000
H1b	Professionalism → Perceived identification	0.155	0.153	0.048	3.224	0.001
H2a	Interaction → Perceived trust	0.202	0.202	0.038	5.323	0.000
H2b	Interaction → Perceived identification	0.215	0.215	0.051	4.213	0.000
H3a	Integrity → Perceived trust	0.236	0.236	0.043	5.521	0.000
H3b	Integrity → Perceived identification	0.165	0.166	0.048	3.464	0.001
H4a	Reputation → Perceived trust	0.186	0.187	0.034	1.344	0.179
H4b	Reputation → Perceived identification	0.046	0.046	0.039	4.812	0.000
H5a	Attractiveness → Perceived trust	0.236	0.237	0.041	5.434	0.000
H5b	Attractiveness → Perceived identification	0.225	0.226	0.052	4.538	0.000
H6	Perceived trust → CPB	0.467	0.468	0.034	13.923	0.000
H7	Perceived identification → CPB	0.425	0.425	0.035	12.063	0.000

. The results reveal that the path coefficient for hypothesis H1a is β = 0.247, T = 5.410, p < 0.001, indicating a significant positive effect of streamer professionalism on perceived trust. Accordingly, hypothesis H1a is supported. Assuming in H1b, β = 0.155, T = 3.224, p = 0.001, this indicates that professional expertise also has a significant positive effect on perceived identification. Thus, H1b is supported. Furthermore, the path coefficients for H2a, H2b, H3a, H3b, H5a, H5b, H6, and H7 are all significantly greater than 0 (T > 1.96, p < 0.05), indicating that the corresponding hypotheses are supported.

Notably, the path coefficient in H4a did not reach significance (T = 1.344, p = 0.179), indicating that reputation does not have a significant effect on perceived identification. This result may be related to frequent trust crisis incidents in the current livestream e-commerce, such as repeated “persona collapses” of prominent streamers, leading to reduced consumer sense of identification with high-profile streamers and reflecting the limitations of reputation in identity construction within specific contexts.

(2) Mediation Effect Test

To examine the mediating roles of perceived trust and perceived identification in the model, this study utilized the bootstrapping sampling method for path analysis and assessed the extent of mediation effects based on the Variance Accounted For (VAF). According to Yang et al. (2019), VAF < 20% indicates no mediation effect, 20% ≤ VAF ≤ 80% indicates partial mediation, and VAF > 80% indicates full mediation [62].

The analysis results are presented in

Table 6. Mediating effect test results.

Hypothesis	Antecedent Variables	Mediating Variables	Outcome Variable	Direct Effects	Indirect Effects	Overall Effects	VAF (%)	Mediating Effect
H8a	Professionalism	Perceived trust	CPB	0.090	0.115	0.205	56.10	Partial mediation
H9a				0.139	0.066		32.20	Partial mediation
H8b	Interaction			0.006	0.094	0.100	94.00	Complete mediation effect
H9b				0.009	0.091		91.00	Complete mediation effect
H8c	Integrity			0.075	0.110	0.185	59.46	Partial mediation
H9c				0.115	0.070		37.84	Partial mediation
H8d	Reputation	Perceived identification		0.083	0.097	0.180	53.89	Partial mediation
H9d				0.159	0.021		11.67	No intermediary effect
H8e	Attractiveness			0.076	0.105	0.181	82.87	Complete mediation effect
H9e				0.081	0.100		55.25	Partial mediation

. In the H8a, the VAF value is 56.10%, representing partial mediation, which indicates that perceived trust partially mediates the relationship between professionalism and CPB. In the H8b, the VAF value is 94.00%, exceeding the 80% threshold, indicating that perceived trust fully mediates the relationship between interaction and CPB. In contrast, in the H9d, the VAF value is only 11.67%, below the 20% standard, indicating that perceived identification does not have a significant mediating effect between reputation and CPB. The results of other path tests also demonstrate that perceived trust and perceived identification exhibit partial or full mediation effects in their respective paths.

5.5. Artificial Neural Network Analysis

ANN is a parallel distributed processor composed of many interconnected processing units, with the capability to store and utilize experiential knowledge [63]. This model can not only identify linear and nonlinear relationships under non-normal distributions, thus enhancing predictive accuracy, but also, owing to its adaptive properties, rapidly respond to changes in data structure, thereby fitting complex and dynamic research environments. These advantages allow it to effectively mitigate the limitations of traditional research methods, enabling precise prediction and importance ranking of key influencing factors. In recent years, ANNs have been extensively applied in consumer behavior research, particularly demonstrating distinct value in addressing complex irrational decision-making problems [2,53,64].

Structurally, a typical ANN comprises the input layer, hidden layers, and an output layer. Among these, the feedforward-backpropagation multilayer perceptron is widely employed to achieve nonlinear discriminant analysis [2,55]. Once the network type is established, the primary task is to appropriately configure the hidden layer structure: each layer consists of multiple neurons connected to the subsequent layer via adaptive synaptic weights. Each neuron multiplies the input signals by their respective weights, sums them, and then transforms the result into an output value through a nonlinear activation function [64]. Secondly, it is necessary to determine the number of neurons in each layer—the number of neurons in the input and output layers corresponds to the number of independent and dependent variables, respectively, which are relatively clear. However, determining the number of neurons in the hidden layer requires careful consideration of multiple factors [65], and the most effective method is dynamic adjustment based on the number of neurons in the input and output layers [66].

The hidden and output layers of the activation function are set to sigmoid, and the number of nodes in the hidden layer is automatically generated. The sigmoid function is common in ANN, calculated as Formula (1):

$$\mathrm{f}\left(\mathrm{x}\right)={\displaystyle \frac{1}{1+e^{-ax}}}, \mathrm{a}>0$$

(1)

where a is the slope parameter of the sigmoid. With the change in a sigmoid with different slopes can be obtained.

The sum square of error (SSE) is calculated by:

$$\mathrm{S}\mathrm{S}\mathrm{E}={\displaystyle \frac{1}{2M}}{\sum }_{m=1}^M{\sum }_{n=1}^N{\left(d_{mn}-o_{mn}\right)}^2$$

(2)

Root mean square error (RMSE) is calculated by:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{t}}}{\sum }_{i=1}^t{\left(y_i-\widehat{y_i}\right)}^2$$

(3)

R Squared (${\mathrm{R}}^2$) is calculated by:

$$R^2=1-{\displaystyle \frac{RMSE}{{s_y}^2}}$$

(4)

Subsequently, this study employed SPSS 24 to perform the ANN analysis on key variables that had passed hypothesis testing [2]. As shown in Figure 3, the study constructed three sub-models with clearly defined input-output structures: Model 1 uses four variables—professionalism, interaction, integrity, and attractiveness—as input layer neurons, and perceived trust as the output layer neuron. Model 2 incorporates reputation based on Model 1, comprising six input layer neurons and using perceived identification as the output. Model 3 employs perceived trust and perceived identification as input layer variables, with CPB as the output layer neuron, thereby establishing a complete path prediction framework.

During model training, a 10-fold cross-validation approach was utilized to prevent overfitting and improve the models’ generalization ability [2,54]. Specifically, 90% of the total sample was allocated for training, with the remaining 10% reserved for testing. Predictive accuracy was assessed by RMSE. The closer this value approaches 0, the better the models’ predictive performance [54].

The RMSE values and goodness of fit for the training and testing models are presented in

Table 7. The RMSE values of ANN models.

Neural Networks	Model 1 (${\mathit{R}}^2=71.62$%)				Model 2 (${\mathit{R}}^2=74.98$%)				Model 3 (${\mathit{R}}^2=70.39$%)
	Training		Testing		Training		Testing		Training		Testing
	SSE	RMSE	SSE	RMSE	SSE	RMSE	SSE	RMSE	SSE	RMSE	SSE	RMSE
ANN1	1.218	0.058	0.241	0.077	1.526	0.064	0.190	0.068	0.825	0.047	0.126	0.055
ANN2	1.992	0.074	0.312	0.087	2.477	0.082	0.467	0.107	0.888	0.049	0.289	0.084
ANN3	1.215	0.058	0.396	0.098	0.999	0.052	0.426	0.102	0.708	0.044	0.492	0.110
ANN4	1.745	0.069	0.242	0.077	0.747	0.045	0.317	0.088	1.041	0.053	0.148	0.060
ANN5	1.294	0.059	0.262	0.080	1.152	0.056	0.373	0.095	0.920	0.050	0.391	0.098
ANN6	1.061	0.054	0.365	0.094	0.731	0.045	0.435	0.103	1.693	0.068	0.129	0.056
ANN7	1.283	0.059	0.274	0.082	1.008	0.052	0.391	0.098	1.073	0.054	0.116	0.053
ANN8	1.217	0.058	0.385	0.097	1.214	0.058	0.360	0.094	1.404	0.062	0.389	0.097
ANN9	1.034	0.053	0.299	0.085	1.061	0.054	0.488	0.109	1.184	0.057	0.380	0.096
ANN10	1.550	0.065	0.222	0.074	0.751	0.045	0.397	0.098	1.208	0.057	0.119	0.054
Average	1.361	0.061	0.300	0.085	1.167	0.055	0.384	0.096	1.094	0.054	0.258	0.076
S.D.	0.308	0.007	0.063	0.009	0.522	0.011	0.085	0.012	0.293	0.007	0.146	0.023

. The RMSE values for the training sets of the three sub-models are 0.061, 0.055, and 0.054, respectively, while those for the testing sets are 0.085, 0.096, and 0.076, respectively. All RMSE values remain low and close to 0, indicating that each ANN demonstrates good predictive accuracy and fitting performance.

To further assess the explanatory power of the models, this study employs the coefficient of determination (R²) to measure the percentage of variance in CPB explained by the ANNs [2]. As shown in Table 7, the predictive accuracies of the three models for CPB are 71.62%, 74.98%, and 70.39%, respectively, indicating that the models overall exhibit good predictive effectiveness and theoretical applicability.

Secondly, to evaluate the predictive strength of the input neurons, a sensitivity analysis was performed. The relative importance of each input neuron was normalized by dividing its score by the maximum relative importance score and expressing it as a percentage [2,55]. As shown in

Table 8. Results of sensitivity analysis.

Neural Network	Model 1				Model 2					Model 3
	Professionalism	Interactivity	Integrity	Attractiveness	Professionalism	Interactivity	Integrity	Attractiveness	Reputation	Perceived Identification	Perceived Trust
ANN1	0.257	0.19	0.297	0.256	0.186	0.212	0.225	0.186	0.19	0.421	0.579
ANN2	0.239	0.272	0.245	0.244	0.202	0.204	0.223	0.157	0.215	0.401	0.599
ANN3	0.228	0.225	0.282	0.265	0.213	0.202	0.202	0.223	0.161	0.455	0.545
ANN4	0.236	0.304	0.238	0.222	0.169	0.167	0.259	0.272	0.133	0.51	0.49
ANN5	0.259	0.269	0.248	0.224	0.14	0.254	0.226	0.216	0.164	0.468	0.532
ANN6	0.243	0.219	0.266	0.272	0.149	0.211	0.238	0.251	0.15	0.499	0.501
ANN7	0.274	0.261	0.24	0.225	0.168	0.245	0.208	0.208	0.17	0.451	0.549
ANN8	0.264	0.172	0.34	0.224	0.188	0.234	0.176	0.251	0.151	0.465	0.535
ANN9	0.228	0.205	0.299	0.268	0.147	0.213	0.263	0.248	0.129	0.452	0.548
ANN10	0.239	0.281	0.271	0.209	0.129	0.175	0.249	0.241	0.206	0.446	0.554
Avg. Importance	0.2467	0.2398	0.2726	0.2409	0.1691	0.2117	0.2269	0.2253	0.1669	0.4568	0.5432
Normalized Importance	89.40%	80.40%	97.80%	100.00%	68.80%	100.00%	85.10%	85.00%	69.50%	83.60%	100.00%

, in Model 1, attractiveness has the greatest impact on perceived trust, followed by integrity (97.8%), professionalism (89.40%), and interactivity (80.4%). In Model 2, interactivity has the greatest impact on perceived identification, followed by integrity (85.10%), attractiveness (85.00%), reputation (69.50%), and professionalism (68.80%). In Model 3, perceived trust has the greatest impact on consumer purchase behavior, followed by perceived identification.

Interestingly, in Model 1, the SEM results indicate that professionalism is the most significant factor influencing perceived trust, followed by integrity, attractiveness, and interaction. In contrast, ANN analysis identifies attractiveness as the primary predictor of perceived trust, followed by integrity, professionalism, and interaction, demonstrating a clear difference in importance rankings between the two methods. In Model 2, SEM analysis reveals that attractiveness exerts the greatest influence on perceived identification, followed by interaction, integrity, professionalism, and reputation. Conversely, ANN analysis shows that interaction is the most influential factor, followed by integrity, attractiveness, reputation, and professionalism, indicating differing variable rankings across the two methods. However, in Model 3, the SEM and ANN results are consistent, both indicating that perceived trust exerts a greater influence on CPB than perceived identification. According to Xiong et al. (2022) [55] and Wang et al. (2025) [2], this discrepancy primarily arises from the ANN’s capacity to capture nonlinear and non-compensatory relationships among variables. Moreover, it generally achieves higher predictive accuracy, allowing a more precise elucidation of the relative importance of factors within complex decision-making processes.

5.6. Main Findings

First, streamers’ integrity, as a core intrinsic moral trait in the central route, exerts a significant positive effect on both consumers’ perceived trust and perceived identification, which is a key finding different from existing research focusing on external streamer traits. Second, streamers’ reputation, as a peripheral route cue, has no significant positive impact on consumers’ perceived trust but can effectively promote the formation of perceived identification, revealing the inconsistent effects of reputation on different consumer psychological variables. Third, all central route traits (professionalism, interactivity, integrity) and peripheral route traits (attractiveness, reputation) have significant positive effects on at least one of the two mediating variables (perceived trust/perceived identification), verifying the rationality of the dual-path classification of streamer traits based on ELM. Fourth, except for the mediating path of reputation–perceived trust–CPB, all other proposed mediating paths are empirically validated, indicating that perceived trust and perceived identification play important mediating roles in the influence of streamer traits on CPB. Fifth, ANN analysis shows that the nonlinear combination of streamer traits has a stronger predictive power for consumer psychological variables and purchase behavior than single individual traits, and the importance ranking of variables identified by ANN is different from that of SEM, which reflects the complex nonlinear interaction among streamer traits.

6. Research Conclusions and Prospects

6.1. Conclusions

Based on the dual-path analysis framework of ELM, this study reveals the differentiated mechanisms through which streamer traits influence CPB. The core conclusions of the study are as follows: (1) The dual-path information processing framework of ELM is applicable to the research of livestream e-commerce consumer behavior: central route traits (professionalism, interactivity, integrity) drive consumers’ rational decision-making by establishing cognitive and emotional trust through in-depth information processing, while peripheral route traits (attractiveness, reputation) promote consumers’ identity-based consumption behavior by eliciting emotional resonance and intuitive perception through heuristic information processing. (2) Integrity is a key intrinsic trait of streamers that affects consumers’ perceived trust and identification in the context of Chinese livestream e-commerce’s quality transformation, which fills the gap of existing research ignoring the influence of streamers’ moral traits. (3) Reputation has an asymmetric effect on consumers’ perceived trust and identification: it cannot effectively enhance perceived trust but can promote perceived identification, which reflects the inherent tension between “traffic (reputation)” and “quality (trust)” in the current livestream e-commerce industry, and also challenges the mainstream view that reputation inevitably improves consumer trust. (4) Perceived trust and perceived identification play important mediating roles in the influence of streamer traits on CPB, and most streamer traits affect consumers’ purchase behavior through the indirect path of “trait perception–psychological variable–purchase behavior”. (5) There are significant nonlinear combination effects among streamer traits, and the optimized combination of multiple traits has a stronger stimulating effect on consumers’ purchase behavior than the independent enhancement of any single trait, which provides a new theoretical basis for the precision matching within the “streamer–product–platform” framework in live streaming e-commerce.

6.2. Theoretical Significance

At the theoretical level, this study advances the research framework of CPB in three ways: first, it transcends the traditional dichotomy between rationality and emotion by explicitly proposing that cognition and emotion simultaneously drive consumer decision-making, thereby deepening the understanding of the formation mechanism of behavioral intentions. Secondly, by introducing a dual-path analysis model, this study systematically elucidates the differentiated influence mechanisms of various information processing pathways on CPB, thereby constructing a more comprehensive theoretical explanatory framework. Finally, from a methodological perspective, this study achieves significant innovation by integrating ANN with traditional SEM, which not only verifies the linear relationships among variables but also reveals complex nonlinear interactions among influencing factors, substantially enhancing the predictive accuracy and explanatory power regarding CPB. This methodological breakthrough provides important theoretical support and methodological reference for the paradigm shift in consumer behavior research from explanation to prediction.

6.3. Practical Significance

(1) Establish a governance mechanism integrating technological empowerment and institutional safeguards to improve the overall credibility of the livestream e-commerce industry. It is recommended that livestream e-commerce platforms develop a multi-level governance framework of “AI monitoring + credit evaluation + diversified co-governance” based on the research conclusion that integrity is the key factor for building consumer trust: employ artificial intelligence and big data technology to achieve real-time monitoring and risk early warning of livestream content, focusing on identifying and preventing false advertising and non-compliant marketing behaviors related to streamers’ integrity; establish a cross-platform mutually recognized streamer credit evaluation system with integrity as the core indicator, and formulate corresponding reward and punishment mechanisms for streamers with different credit levels; build a comprehensive governance ecosystem guided by government leadership, platform main responsibility, industry self-regulation, and social participation by establishing convenient consumer complaint channels and introducing social supervisors, so as to create a trustworthy livestream shopping environment for consumers.

(2) Develop a streamer cultivation and evaluation system centered on integrity, and reverse the industry’s “traffic-centric” logic through incentive mechanisms. Based on the research finding that integrity significantly enhances consumer trust and identification, livestream platforms and MCN institutions should incorporate streamer integrity as a core criterion in talent selection, training, and career development, and promote the establishment of a dual-core competency model for streamers comprising professional competence and moral quality. Specifically, professional credit files and dynamic evaluation mechanisms can be introduced for streamers, and the integrity performance of streamers can be linked to resource allocation, such as traffic recommendation and promotional opportunities, as well as commission structure adjustment, to guide streamers to pay attention to the cultivation of moral quality and form a good industry atmosphere of “integrity-based”. Specifically, “integrity” in the digital commerce context should be operationalized into measurable multi-dimensional indicators, including:

Information authenticity: the proportion of product information (functions, ingredients, prices) that is verified as true by platform audits or consumer feedback, and the frequency of false advertising violations;

Commitment fulfillment rate: the proportion of commitments (e.g., after-sales service, discount validity, product quality guarantees) fulfilled by streamers after the live broadcast, which can be tracked through post-purchase consumer satisfaction surveys and after-sales dispute data;

Consumer feedback indicators: the streamer’s product return rate (compared with the industry average for the same category), the proportion of complaints related to “false propaganda” or “breach of commitment” among total complaints, and the positive word-of-mouth rate related to integrity in consumer comments;

Compliance records: whether the streamer has a history of administrative penalties for false advertising, unfair competition, or other integrity-related violations.

Platforms can establish professional credit files for streamers, integrate the above indicators into a dynamic integrity scoring system (e.g., 0–100 points, updated monthly), and directly link the integrity score to resource allocation: streamers with high integrity scores (e.g., ≥85 points) are given priority in traffic recommendation, participation in platform-level promotional activities, and reduced commission rates; those with low scores (e.g., <60 points) are restricted in live broadcast frequency, required to participate in integrity training, and even removed from the platform if violations are serious. This way, the abstract concept of integrity is transformed into specific, operable platform management metrics, providing clear guidance for industry practice. Specifically, professional credit files and dynamic evaluation mechanisms can be introduced for streamers, and the integrity performance of streamers can be linked to resource allocation, such as traffic recommendation and promotional opportunities, as well as commission structure adjustment, to guide streamers to pay attention to the cultivation of moral quality and form a good industry atmosphere of “integrity-based”.

(3) Deepen content and scenario innovation based on the dual-path theory of ELM, and meet consumers’ dual demands for cognition and emotion. Based on the research conclusion that central and peripheral routes drive consumer behavior through different mechanisms, livestream operation practitioners should construct an immersive livestream scenario of “professional content + emotional connection”: on the central route, strengthen the streamer’s professional product knowledge reserve and transparent disclosure of product information, improve the quality of rational persuasion, and meet consumers’ cognitive needs for product information; on the peripheral path, enhance user engagement and emotional identification through diversified interactive design and emotional storytelling, give play to the role of emotional resonance, and realize the synergistic effect of rational persuasion and emotional resonance in livestream marketing.

6.4. Research Limitations and Prospects

This study has certain limitations in sample representativeness, variable systems, and research methods, which constrain the depth and scope of the conclusions. which constrain the depth and scope of the conclusions, and also point out the direction for future research. First, it is necessary to clarify the boundary conditions of the research model: this study is conducted in the context of Chinese livestream e-commerce, and the research conclusions are mainly applicable to the cultural background of collectivism in China and mainstream livestream e-commerce platforms with real-person streamers as the main body; the sample is mainly concentrated in the young and middle-aged consumer groups in eastern China with a bachelor’s degree and middle-to-low income, so the model may have different applicability in other cultural backgrounds (e.g., individualism), other types of livestream platforms (e.g., cross-border livestream platforms, virtual streamer platforms), and other consumer groups (e.g., high-income groups, elderly groups). The transferability of the research model needs to be further verified by subsequent cross-cultural and cross-platform comparative research.

Second, there is a notable “young and low-income group bias” in the sample, which may have a potential impact on the research findings. Specifically, the sample includes a large proportion of young respondents (18–30 years old accounting for 62.01%) with undergraduate education (65.93%) and a monthly income of 2000–4000 yuan (49.75%), many of whom are students or new employees. This demographic group typically has distinct consumption characteristics: they are more familiar with digital media, pursue social interaction and emotional resonance in livestream shopping, and are more sensitive to streamers’ “interactivity” (e.g., real-time responses, personalized communication) and “attractiveness” (e.g., appearance, charisma). Therefore, their evaluations of these two traits may be over-indexed compared to older (40 years old and above) and high-income (monthly income above 8000 yuan) consumer groups. In contrast, older and high-income consumers often have more mature consumption concepts, pay more attention to the practical value and quality of products, and may prioritize streamers’ “professionalism” (e.g., product expertise, industry experience) over interactivity and attractiveness when making purchase decisions. This sample bias may lead to the overestimation of the influence of interactivity and attractiveness in the research model, while the influence of professionalism may be underestimated. Therefore, the research conclusions should be cautiously generalized to broader consumer groups, especially older and high-income segments. Future research should adopt stratified sampling to include more respondents from different age and income groups, and conduct multi-group comparative analysis to verify whether the influence mechanism of streamer traits on consumer behavior varies across demographic groups, thereby enhancing the generalizability and robustness of the research findings.

Specifically, the sample shows a concentration in demographic characteristics, which may affect the generalizability of the findings. Studies on streamer traits mainly focus on common dimensions, without adequately incorporating contextualized traits such as humor and audience control ability. Methodologically, cross-sectional design and self-reported data are limited in capturing the dynamic processes of consumer decision-making and the underlying physiological mechanisms.

To address these limitations, future research may adopt the following approaches: in data collection, implement stratified sampling and longitudinal surveys to acquire more representative and dynamically evolving panel data. In the design of variables, develop a more comprehensive framework of streamer traits and explore their synergistic or substitution effects. Regarding research methods, combining neuroscientific experiments (such as eye-tracking and electroencephalogram) with hybrid modeling techniques can empirically validate and extend the intrinsic mechanisms of the dual-pathway theoretical model from multiple dimensions.