The study explores the causal framework of the major technical and user-oriented variables that define the efficacy of e-commerce recommendation systems (RSs) that are based on artificial intelligence (AI). Although the interdependency between algorithmic quality attributes (e.g. accuracy, diversity) and user perception variables (e.g. satisfaction, trust) has been studied in the research literature before, their directional relationships have not been well explored. In order to fill this gap, the paper uses the Decision-Making Trial and Evaluation Laboratory (DeMaTel) to simulate the cause and effect relationships among seven factors, including accuracy, completeness, diversity, and novelty, ease of use, satisfaction, and trust. The data were the pairwise influence data of 15 domain experts in academia and industry and processed through a standardized DeMaTel procedure. The findings establish accuracy, completeness, diversity, novelty, and ease of use as causal factors, and satisfaction and trust as dependent factors with high degrees of prominence. Most of the net causal influence is on accuracy, so it can be seen as the main determinant of the system’s performance and perceptions of the users. The combined application of both technical and behavioural constructs in a unified causal model enables the study to offer a new hierarchical perspective of the effectiveness of an RS that goes beyond the traditional correlation-based models. The managerial implications of the findings indicate that upstream technical quality improvement, especially accuracy of recommendations and completeness of information, indirectly leads to user satisfaction and trust, and there is a specific design consideration for AI engineers and e-commerce managers. The contribution of the study to methodology is that it shows how DeMaTel is appropriate in modelling complex socio-technical interactions in the case of recommender systems, and a theoretical contribution is providing a route of causation between the performance of an algorithm and the user-level consequences. Limitations related to expert sample size are acknowledged, and future research directions are proposed.
Mahabir et al. (Tue,) studied this question.