Inherent Preferences or Conformity Effects? A Psychological Dynamics-Aware Causal Recommendation Approach

 

内容摘要:

Recommender systems on e-commerce platforms typically leverage consumers’ behavior sequences to suggest products aligned with their interests. However, disentangling the combined effects of inherent preferences and conformity driven by product popularity on consumer behavior is crucial for addressing personalized demands. Existing causal recommendation methods primarily rely on static disentanglement, which fails to capture the dynamic evolution of causes and especially the role of psychological dynamics in shaping behavioral mechanisms. Drawing on the theories of marketing funnel and consumer behavioral biases, this study proposes STAGE (State Transition-Aware Generative model for cause disEntanglement), a novel dynamic causal recommendation approach that models the evolution of latent causes and the impact of psychological dynamics throughout consumer shopping journeys. STAGE employs causal embedding to generate representations of inherent preference and conformity effect, formulates psychological dynamics using state transition paths to characterize differentiated behavioral mechanisms, incorporates the effect of accumulated shopping experiences on subsequent journeys, and utilizes proxy variables to achieve cause disentanglement and model identifiability. Multi-view experiments on real-world datasets demonstrate that STAGE achieves superior performance in both in-sample next-item prediction and inherent preference-oriented recommendation, while effectively disentangling the underlying causes of consumer behavior for targeted marketing. This study contributes to the computational design science realm by designing a psychological dynamics-aware cause disentanglement approach and revealing consumer behavioral mechanisms under different psychological paths.