Purpose Most studies on algorithm aversion are scenario-based online or laboratory experiments and review type analysis, and few studies apply empirical research methods. Further to this, the research is mainly focused on developed countries, with nearly two-thirds of algorithm aversion and appreciation surveys and experimental studies conducted using American samples. More than 80% of the studies used European or United States participants. However, research in the context of developing countries (e.g. China) may provide different insights. Design/methodology/approach Different from traditional empirical research methods, this study combines SEM and ANN two-stage method, which does not only cover the hypothesis testing of the linear relationship in the compensation model, but also captures the nonlinear non-compensation relationship in the neural network model. Findings Algorithm transparency, algorithm accuracy, tolerance and surprise degree have a significant negative impact on algorithm aversion, while user perceived algorithm control and expectancy violations have a significant positive impact on algorithm aversion. Algorithm transparency, algorithm accuracy and tolerance significantly negatively impact expectancy violations, while overconfidence significantly positively impacts expectancy violations. Algorithm accuracy significantly positively impacts surprise degree, while overconfidence significantly negatively impacts surprise degree. Timeliness not only has a significant negative impact on algorithm aversion, but also plays a significant moderating role on the impact of expectancy violations and surprise degree on algorithm aversion. In terms of the importance of standardization, expectancy violations are the most important, followed by algorithm accuracy, tolerance, surprise degree, user perceived algorithm control, algorithm transparency, and timeliness. The ANN model can predict algorithm aversion with an accuracy of 85.4%. Originality/value This study combines SEM and ANN two-stage method was used to successfully verify the effects of algorithm characteristics, receiver characteristics, recommendation system characteristics and situational factors (expectancy violations, surprise degree) on algorithm aversion. It provides beneficial practical insights for recommendation systems, algorithm developers to optimize algorithms, and users to effectively utilize algorithms.
Wang et al. (Wed,) studied this question.