Research on the formation mechanism of users’ algorithm aversion behavior in the context of intelligent recommendations: an empirical analysis based on the FATE framework and SEM-ANN method

While enjoying personalized services, users are concerned about potential threats, such as information cocoons and privacy leaks, arising from the widespread use of intelligent recommendation algorithms. Therefore, from the algorithmic technology dimension, this study introduces the Fairness, Accountability, Transparency, and Explainability (FATE) framework—representing users’ perception of algorithmic technical attributes—into the cognition-affect-conation (CAC) model pathway, and constructs a chain mediation model with technology perception as the antecedent and psychological mediation as the mechanism. A questionnaire survey was conducted; structural equation modeling (SEM) was used to test the hypothesized paths, and artificial neural network (ANN) analysis was incorporated as a supplementary analytical tool to capture nonlinear relationships among key variables. The results show that users’ negative perceptions of algorithmic technical attributes are positively associated with their behavioral intention toward algorithm aversion, with perceived algorithmic opacity exhibiting the strongest association within the current sample. Perceived risk and perceived manipulative intent play parallel mediating roles between technical perception and aversion behavior, and a chain mediation effect is also present. Self-efficacy negatively moderates the path from perceived risk to behavioral intention. At the theoretical level, by integrating the FATE framework (technical attribute perception) with the CAC psychological pathway and adopting a hybrid structural equation modeling-artificial neural network (SEM-ANN) quantitative approach, this study offers a complementary explanatory framework for algorithm aversion. At the practical level, it suggests that platform designers may focus on improving algorithmic transparency and fairness, establishing a user control system, leveraging self-efficacy to mitigate algorithm aversion, and adopting precision intervention strategies based on the observed nonlinear effects, thereby providing a practical reference for optimizing algorithm design and enhancing user trust.