Research on the Impact of Personalized Recommendation on Consumers Purchase Intention on Online Shopping Platforms: A Multi-Dimensional Perceived Value Perspective

1. Introduction

The digital commerce landscape has undergone a fundamental transformation driven by the integration of machine learning and big data analytics. Online shopping platforms now deploy sophisticated personalized recommendation systems (PRS) as core strategic tools to manage information overload, reduce consumer search costs, and guide purchasing decisions (Zhang et al., 2022). In markets characterized by intense competition—such as China's e-commerce ecosystem encompassing Taobao, JD.com, Pinduoduo, and Douyin—the ability to deliver highly relevant, individualized product suggestions has become a decisive competitive advantage.

Despite the widespread deployment of PRS, a significant theoretical gap persists: existing research has largely treated personalized recommendation as a monolithic construct, failing to account for the fundamentally different psychological mechanisms activated by distinct algorithmic logics. Collaborative Filtering (CF) leverages social proof, Content-Based Filtering (CBF) reinforces individual taste, and Hybrid systems seek to combine these approaches while incorporating contextual signals. Each type is hypothesized to generate distinct patterns of perceived value in the consumer, yet comparative empirical analysis remains scarce.

This paper addresses this gap by applying the SOR model (Mehrabian & Russell, 1974) to propose a novel Typology-Value-Behavior (TVB) framework. The recommendation type constitutes the external Stimulus (S); multi-dimensional perceived value constitutes the internal Organism (O); and purchase intention constitutes the behavioral Response (R). The paper is structured as follows: Section 2 reviews theoretical foundations; Section 3 presents the research gap analysis and literature comparison; Section 4 develops the research model, measurement instruments, and hypotheses; Section 5 outlines the research methodology; Section 6 discusses expected contributions; and Section 7 concludes with implications and future directions.

2. Theoretical Foundations

2.1 Personalized Recommendation Systems: Typology and Mechanisms

Personalized recommendation is formally defined as an information filtering technology that predicts a user's preference for items by leveraging historical behavioral data, user profiles, and contextual information (Zhang et al., 2022). In e-commerce, the primary function of PRS is to mitigate information overload and reduce consumer decision effort (Chen et al., 2022). Three principal algorithmic types are distinguished, each with distinct operational logic and consumer-facing characteristics, as summarized in Table 5 later in this paper.

Collaborative Filtering (CF) recommends items based on the aggregated preferences of users with similar behavioral profiles, often manifested as 'Customers who bought this also bought...' suggestions. CF operates on social proof and is particularly effective at surfacing cross-category associations (Wu et al., 2021). Content-Based Filtering (CBF) recommends items with attributes similar to those previously interacted with by the individual user, reinforcing personal taste and generating high perceived relevance (Kang et al., 2022). Hybrid systems combine CF, CBF, and real-time contextual signals to improve both accuracy and serendipity (Yuan et al., 2023).

Two significant challenges constrain PRS effectiveness: the personalization-privacy paradox, wherein heightened personalization triggers consumer discomfort (Aksoy et al., 2022), and the filter bubble effect, wherein narrow algorithmic confinement limits product exposure (Ge et al., 2020). These challenges underscore the need to understand how recommendation type shapes consumer psychology, not just technical performance.

2.2 Perceived Value: A Multi-Dimensional Mediating Construct

Perceived value is conceptualized as a consumer's overall assessment of utility calculated as perceived benefits relative to perceived costs (Zeithaml, 1988). Building on the PERVAL scale (Sweeney & Soutar, 2001), this study conceptualizes four value dimensions. Functional Value refers to core utility and efficiency derived from accurate recommendations (Venkatesh et al., 2012). Emotional Value encompasses hedonic affective states—enjoyment and curiosity—generated by the discovery experience (Roy et al., 2020). Social Value pertains to the recommendation's capacity to enhance social self-concept (Zhang & Benyoucef, 2016). Novelty Value captures the pleasure derived from encountering unexpected, serendipitous suggestions (Jiang et al., 2021).

2.3 Purchase Intention in Algorithmic Environments

Purchase intention represents a consumer's cognitive commitment to undertake a purchasing behavior and serves as the strongest proximal predictor of actual buying conduct online (Pavlou & Fygenson, 2016). Recommendations perceived as inaccurate, intrusive, or manipulative can elicit psychological reactance, suppressing purchase intention (Li et al., 2021). This non-linearity underscores the importance of the organism-level mediation: it is not the recommendation itself, but the consumer's internal value assessment, that ultimately drives behavioral response.

2.4 The SOR Framework as Integrative Architecture

The Stimulus-Organism-Response (SOR) model (Mehrabian & Russell, 1974) explains how platform-level stimuli influence behavioral outcomes through the mediation of internal cognitive and affective states (Li & Ku, 2022). Applied here: PRS type is the external stimulus; multi-dimensional perceived value is the organism's internal processing; and purchase intention is the behavioral response. Table 1 presents the proposed TVB framework built on this architecture.

Source: Author's own construction based on Mehrabian & Russell (1974) and Sweeney & Soutar (2001)

3. Literature Gap Analysis

A systematic review of the extant literature reveals three persistent gaps that the present study is designed to address. First, the majority of prior studies treat PRS as a generic stimulus without distinguishing between the fundamentally different psychological logics of CF, CBF, and Hybrid systems. Second, perceived value, when modeled, is typically reduced to a single dimension—most often functional value—neglecting the emotional, social, and novelty dimensions increasingly salient in China's discovery-driven commerce landscape. Third, mediation analysis in this domain has been limited to single-pathway models, failing to capture the simultaneous and comparative influence of multiple value dimensions.

Table 2 presents a structured comparison of five representative prior studies against the present study, mapping each work's treatment of the PRS stimulus, value dimensions examined, theoretical framework employed, and the specific gap it leaves unaddressed.

Source: Author's synthesis based on systematic literature review (2019–2024)

As Table 2 demonstrates, no prior study has simultaneously (a) distinguished between CF, CBF, and Hybrid recommendation types as separate stimuli, (b) modeled all four perceived value dimensions as parallel mediators, and (c) tested differential mediation pathways within an SOR framework calibrated for China's mobile-first e-commerce context. The present study is designed to fill precisely this gap through the proposed TVB framework.

4. Research Model, Measurement, And Hypotheses

4.1 Measurement Instruments

All constructs are operationalized using established, validated scales adapted for the Chinese digital commerce context. Items use a seven-point Likert scale (1 = Strongly Disagree; 7 = Strongly Agree). The independent variable—Personalized Recommendation Type—is measured using scenario-based items presenting distinct algorithmic descriptions. This scenario-based approach overcomes the well-documented challenge of measuring abstract algorithmic logics in survey research. Table 3 presents the full measurement instrument with sample items and their theoretical sources.

Note: All scales adapted from cited sources for Chinese e-commerce context. 7-point Likert scale used throughout.

4.2 Research Hypotheses

Based on the theoretical foundations reviewed in Section 2 and the gaps identified in Section 3, this study proposes 13 hypotheses organized across four groups: (H1) direct effects of recommendation types on purchase intention; (H2) differential effects of recommendation types on perceived value dimensions; (H3) effects of perceived value dimensions on purchase intention; and (H4) mediating role of perceived value. Table 4 presents the complete hypothesis set with theoretical justification for each.

Note: H = Hypothesis; CF = Collaborative Filtering; CBF = Content-Based Filtering; FV = Functional Value; EV = Emotional Value; SV = Social Value; NV = Novelty Value.

5. Research Methodology

5.1 Research Design

This study employs a quantitative cross-sectional survey design, appropriate for testing theoretically derived hypotheses regarding latent psychological constructs in large consumer samples (Hair et al., 2019). The primary data collection instrument is a structured self-administered questionnaire deployed through Chinese online survey platforms (Wenjuanxing and Credamo), targeting active online shoppers with experience across major Chinese e-commerce platforms. Table 5 provides a detailed comparison of the three recommendation types to be operationalized as stimuli in the survey instrument.

Source: Author's synthesis based on Wu et al. (2021), Kang et al. (2022), and Yuan et al. (2023)

5.2 Sample Size and Sampling Strategy

The required sample size was determined through a priori statistical power analysis using G*Power 3.1. For multiple regression with nine predictors, assuming medium effect size (f² = 0.15), significance level α = 0.05, and target power of 0.80, the minimum required sample is 166 valid responses. To account for the complexity of the parallel mediation model (four mediators) and anticipated attrition, a target of 350 valid responses is established. This exceeds the threshold recommended for bootstrapping-based mediation analysis (Hayes, 2018; Faul et al., 2007). Stratified purposive sampling across diverse online communities ensures representativeness across age cohorts and platform preferences.

5.3 Data Analysis Strategy

Data analysis proceeds in five stages using SPSS 27.0 and the PROCESS macro (Hayes, 2018). Stage 1: descriptive statistics and data screening. Stage 2: measurement model validation through reliability analysis (Cronbach's α > 0.70; CR > 0.70) and validity assessment via Exploratory and Confirmatory Factor Analysis (AVE > 0.50). Stage 3: Pearson correlation analysis. Stage 4: hierarchical multiple regression to test direct effect hypotheses (H1–H3). Stage 5: Hayes' PROCESS Model 4 with 5,000 bootstrap samples and 95% bias-corrected confidence intervals to test parallel mediation (H4). Common method bias will be assessed using Harman's single-factor test and procedural remedies including item randomization and respondent anonymity.

6. Expected Contributions

6.1 Theoretical Contributions

This study advances the SOR framework's application in AI-mediated commerce by introducing granular algorithmic typology as the stimulus construct. This refinement generates greater predictive specificity and enables differential hypothesis testing across recommendation types. The study reconceptualizes perceived value as a multi-dimensional mediating organism-level state in the context of algorithmic service evaluation, explicitly modeling novelty value as a distinct dimension. The proposed TVB framework bridges the theoretical divide between Information Systems research and Marketing/Consumer Behavior research, fostering interdisciplinary dialogue neither field has fully achieved in isolation.

6.2 Practical Contributions

The study generates actionable outputs for e-commerce practitioners in China's competitive digital marketplace. An Algorithm-to-Value Mapping tool will link specific recommendation types to the perceived value dimensions they most effectively enhance, enabling strategic algorithm selection based on desired consumer outcomes. Findings will also inform UX and communication design—specifically how recommendation labels ('Because customers like you chose this' vs. 'Because you viewed this') can amplify intended value perceptions and mitigate privacy-related reactance, particularly relevant under China's evolving data protection frameworks.

6.3 Methodological Contributions

The scenario-based operationalization of algorithmic types addresses the challenge of measuring abstract algorithmic logics in survey research. The parallel mediation design, testing four simultaneous mediators with bootstrapping inference, provides a rigorous methodological template for future research on multi-pathway psychological models in digital consumer behavior contexts.

7. Conclusion

This paper has proposed a theoretically grounded framework for understanding how personalized recommendation types influence consumer purchase intention through the mediating mechanism of multi-dimensional perceived value. By integrating the SOR model with a refined typology of recommendation algorithms and a multi-faceted conceptualization of perceived value, the proposed TVB framework offers both theoretical novelty and practical applicability within the context of China's dynamic e-commerce environment.

The study addresses three critical gaps: the oversimplification of the recommendation stimulus; the incomplete examination of multi-dimensional value mediation; and the need for contextually specific investigation within China's mobile-first ecosystem. The methodology—anchored in a G*Power-validated sample of 350 respondents, scenario-based measurement, and advanced parallel mediation analysis—is designed to generate rigorous and externally valid findings.

Future research should extend the TVB framework through longitudinal designs, examine moderating variables such as consumer innovativeness and platform trust, and conduct cross-platform comparisons between Chinese platforms (Taobao, JD.com, Pinduoduo, Douyin) and Western e-commerce ecosystems to test the generalizability of the proposed model. As recommendation systems evolve toward increasingly sophisticated AI-driven personalization, the theoretical and empirical tools developed in this study provide a foundation for sustained scholarly inquiry into the human-algorithm interface in digital commerce.