BỘ GIÁO DỤC VÀ ĐÀO TẠO ĐẠI HỌC KINH TÉ THÀNH PHÓ HÒ CHÍ MINH BÁO CÁO TỐNG KẾT ĐÈ TÀI NGHIÊN cứu KHOA HỌC THAM GIA XÉT GIẢI THƯỞNG ‘’NHÀ NGHIÊN cứu TRẺ UEH” NĂM 2024 INVESTIGATING CONSUMERS’ RESISTANT REACTIONS TO AI-BASED CONTENT RECOMMENDATION ON SHORT-VIDEO PLATFORMS: A STUDY OF GREEDY AND BIAS RECOMMENDATIONS Thuộc nhóm chuyên ngành : 1 TP. Hồ Chí Minh, tháng 02/2024 ABSTRACT With the popularity of personalized recommendation systems (PRS) such as Facebook, Amazon, Tik Tok,. AI-based recommendation algorithms have received widespread attention from even academia. However, research has provided limited insights into the dark side of AI-based recommendation algorithms, and the underlying mechanisms through which these effects impact psychological and behavioral responses, especially affecting purchasing behavior.
Based on the stressor-strain-outcome (SSO) framework, this study aims to analyze the proposed feature of “greedy" and “bias" that cause associated stressors to information, thereby examining its impact on users' negative psychological and behavioral responses. This research collected 473 online responses and conducted empirical analysis. The results show that both greedy recommendation and bias recommendation algorithms cause information narrowing, information redundancy, information overload, technology intrusiveness, and information disclosure concerns. These stressors can cause negative psychological and behavioral responses in users, which will ultimately influence purchase discontinuation behavior on short video platforms.
The findings of this study contribute to the comprehension of the dark side of AI recommendation algorithms and provide practical suggestions for short-form video application providers. Both theoretical and practical implications are discussed in detail. Keywords: Greedy Recommendation - Bias Recommendation - sso framework Information characteristics - Discontinuous Purchase Behaviour - Dark side ofAI TABLE OF CONTENT ABSTRACT. 2 TABLE OF CONTENT.
3 LIST OF TABLE.5 LIST OF FIGURE. 5 LIST OF ABBREVIATIONS. Necessity of the research. 4 CHAPTER 2: LITERATURE REVIEW.
Stressor-Strain-Outcome Framework. Short-video platforms and AI recommendation algorithm. Discontinuous purchase behavior. Information disclosure concerns.
28 CHAPTER 4: RESEARCH RESULTS. Common method bias (CMB). Limitations and future research. 1 APPENDIX 1: QUANTITATIVE QUESTIONNAIRE.
21 LIST OF TABLE Table 1. The reliability and convergent validity. Heterotrait-monotrait ratio (HTMT). Fornell-Larcker criterion.
Latent variable correlations. Path coefficients - mean, STDEV, T-Values, P-Values.42 LIST OF FIGURE Figure 1: Research model.24 LIST OF ABBREVIATIONS Abbreviation symbols Meaning à Anxiety AI Artificial intelligence AN Anger AVE Average variance extracted BR Bias recommendation CA Cronbach's alpha CMB Common-method bias CR Composite reliability DPB Discontinuous purchase behavior DT Distrust GR Greedy recommendation I DC Information disclosure concerns IN Information narrowing IO Information overload IT Information technology IR Information redundancy IS Information system ISA Internet social anxiety NI Information technology nWOM Negative word of mouth PLS-SEM Partial least squares structural equation modeling PRS Personalized recommendation systems SNS Social networking service sso Stressor-strain-outcome TI Technology intrusiveness 1 CHAPTER 1: INTRODUCTION 1. Research situation In recent years, companies strive to develop and produce more personalized recommendation systems (PRS) and personalized services for their customers to meet user needs. PRS is widely used in short-form video applications (Apps) including TikTok, YouTube Shorts, Instagram Reels, and Facebook Watch.
Short video platforms give users the ability to quickly and easily create and upload videos shorter than five minutes to share with their friends, family, or the entire Internet (Kaye et al. It satisfies users with a variety of information tailored to their interests based on AI recommendation algorithms, in which greedy recommendation and bias recommendation play important roles. Greedy recommendation refers to a system that recommends content with the highest predicted relevance without exploring other potential user preferences (Xing et al. Bias recommendations refer to content that is unfairly biased (O'Neil, 2016).
Although AI recommendation algorithms have potential negative impacts, relatively few studies have systematically investigated the dark side of AI recommendation algorithms from an informational perspective by specifically studying greedy recommendation and bias recommendation. Besides, understanding the stressors caused by greedy recommendation and bias recommendation is an important step to diagnose shortcomings in current strategics and improve user experience and recommendation performance (Ma eỉ al. Furthermore, it is unclear whether the negative effects caused by stressors by the proposed algorithm will affect negative psychological and behavioral reactions, especially discontinuous purchase behavior in users. Therefore, the dark sides of recommendation algorithms should not be overlooked, especially in depth research on the impact of greedy recommendation and bias recommendation on short video platforms.
Previous research has shown that because PRS relies heavily on data collection of users' personal information (e., personally identifiable information and biographical information), users may perceive privacy concerns and threats due to their personal data being covertly collected, which has the potential to create negative reactions in individuals (Chen et al., 2019a, b; Newell and Marabelli, 2015). In the context of online shopping, overload (e., information overload) causes negative emotions (stress) in 2 users, such as anxiety about the website, leading to attitudinal changes, such as subjective states for purchasing decisions (Ding et al. Necessity of the research Several research gaps still remain, despite the fact that the suggested algorithms have drawn attention from researchers due to their broad applicability (Ghasemaghaei et al. First, most studies focus on the positive aspects of recommendation algorithms, such as meeting users' personalized needs and encouraging participation (Liang et al., 2006), while a few focus on the dark side.
Besides, most research articles on the dark side only focus on the effect on negative reactions in users in general, without knowing what that reaction is, most will talk about negative reactions in users “exhausted'’ by users with AI recommendation algorithms. Second, while researchers have found a number of predisposing factors (such as vulnerability and privacy concerns) for user reactions (Aguirre et al., 2015; Chen et al., 2019a, b), they have not given much thought to the function of information. The primary goal of the suggested algorithm, according to Liu et al. (2011), is to satisfy the user's information needs.
It can also respond when certain information characteristics fail to do so or even compromise the user's information needs responses as sources of information stress, leading to unfavorable user reactions. Consequently, conducting extensive research on the function of information in particular is crucial. Third, prior research (Aljukhadar eỉ al., 2012; Benbasat and Wang, 2005) simply looked at suggestion as a broad concept without identifying the particular traits of recommendation algorithms. One of the most noticeable characteristics of the suggested algorithm, for instance, is its "greedy,” which refers to the method by which it uses existing estimations without further investigation (Bastani et al.
Furthermore, the implications of bias recommendation algorithms are rarely discussed in research studies. In order to understand user reaction patterns, it is important to examine the features of greedy and bias recommendation algorithms, which have the capacity to alter consumers' information consumption patterns and create information stressors. This paper examines the negative aspects of recommendation algorithms and clarifies the connection between recommendation features and their impact in light of these research gaps. This research brings new contributions to AI recommendation algorithms on short video platforms.
First, unlike previous studies that focus on research on recommendation 3 algorithms in general, we clearly point out the "greedy" and especially "bias" features of the recommendation algorithm, thereby clarifying its effects on negative user reactions from a stressor perspective. Second, the study adds new knowledge about stressors from the perspective of novel information characteristics including information narrowing, information redundancy, information overload, technology intrusiveness, and information disclosure concerns. In particular, our study is one of the first to identify the influence of information characteristics on user responses to recommendation algorithms. The research results provide future research with a new perspective on information-related factors when studying the dark side of recommendation algorithms.
Third, unlike previous studies that focus on the bright side of recommendation algorithms (Liang et al., 2006), we expand our research perspective to the dark side in a more specific way. Specifically, we investigate users' negative reactions to recommendation algorithms through psychological (anger and anxiety) and behavioral responses in users, specifically focusing on discontinuous purchase behavior from users. This study highlights the weakness of recommendation algorithm research with respect to the dark side of business research. Finally, by building on the sso framework, we advance the understanding of the mechanism of the recommendation algorithm's influence on negative user feedback through the lens of sso.
This framework has been widely applied in the context of information technology to explain negative user reactions from the perspective of stressors (Cao et al. This study also extends the sso framework into the recommendation context, addressing the shortcomings of context-specific factors in the sso literature. General objective We investigate consumers’ resistant reactions to Al-based content recommendations on short-video platforms using greedy and bias recommendations. In addition, the research also strengthens the theoretical framework, develops suggestions for scholars, and provides solutions for marketers, managers, and businesses.
Specific objectives To achieve the above general goal, we address the following specific goals: First, determine how greedy and bias recommendations influence purchase discontinuation behavior through information tensions; 4 Second, the research helps measure and evaluate the recommendation algorithm influencing user behavior in Vietnam; Third, we offer some suggestions and recommendations for retailers to reduce customer stress and provide programs to attract customers when purchasing using short video platforms. Research subjects Customers using short video platforms to purchase in Vietnam. Research methodology Quantitative research: Collect and analyze survey data and test models. Our team surveyed a sample size of 473 short video platform users in Vietnam.
This study is the initial research model testing phase, including determining the sample size, designing a questionnaire based on calculations, and testing the questionnaire with respondents. The next stage was to collect data using targeted random sampling and questionnaire interviews. Finally, the study results were extracted using PLS regression analysis. Research scope Time: 24/09/2023 -15/02/2024 Space: Consumers in Vietnam 5 CHAPTER 2: LITERATURE REVIEW 2.
Stressor-Strain-Outcome Framework There are three parts to the stressor-strain-outcome (SSO) structure. The stressors are environmental stimuli (objective events) that a person encounters that the actor perceives and interprets as troublesome and potentially disruptive. Strain refers to the disruptive impact of stressors on individuals' psychology and emotions, and outcome refers to behavioral or psychological responses to or consequences of strain (Koeske and Koeske, 1993). The underlying principles of the sso framework use a stress perspective to explain environmental stimuli and corresponding outcomes at the individual level (Lazarus, 1966; Lazarus and Folkman, 1984), so the sso framework is relevant to much of the research on the context of technology, especially in research on the dark side of information system use.
For instance, in the context of social media, the fear of missing out, excessive use of social media, and overload are the stressors affecting individuals' emotions and attitudes (strain, such as exhaustion, regret, or dissatisfaction) towards social media, which in turn lead to various adverse outcomes, such as decreased job or academic performance and discontinuous social media usage intentions (Cao et al., 2018; Dhir et al., 2018; Nawaz et al., 2018; Yu et al. In the context of SNS use, scholars have used sso to investigate users' learning performance. SNS use causes technological stresses (stressors), such as overuse, privacy concerns, and self-disclosure, causing psychological stress, which in turn leads to reduced academic performance (outcome) (Cao et al., 2018; Dhir et al. Technostress is defined as individuals' psychological states of mind caused by their inability to cope with the current demands induced by their social media usage (Steelman and Soror, 2017).
Under this stress, users are likely to experience emotional decline, such as psychological fatigue, mental weaknesses, and satisfaction decay (Hsiao, 2017). Furthermore, strain in the form of fatigue or exhaustion results in different negative psychological or physical outcomes. Therefore, the sso framework fits perfectly with the study's primary objective, i., to investigate which factors induce anger and anxiety emotions by the Al recommendation algorithm on a short video platform and its results.