BACF: Bayesian Attentional Collaborative Filtering

The scarcity of explicit feedback data is a major challenge in the design of recommender systems. Although such data are of a high quality due to users’ voluntary provision of numerical ratings, collecting a sufficient amount in real-world service environments is typically infeasible. As an alternative, implicit feedback data are extensively used. However, because implicit feedback represents observable user actions rather than direct preference statements, it inherently suffers from ambiguity as a signal of true user preference. To address this issue, this study reinterprets the ambiguity of implicit feedback signals as a problem of epistemic uncertainty regarding user preferences and proposes a latent factor model that incorporates this uncertainty within a Bayesian framework. Specifically, the behavioral vector of a user, which is learned from implicit feedback, is restructured within the embedding space using attention mechanisms applied to the user’s interaction history, forming an implicit preference representation. Similarly, item feature vectors are reinterpreted in the context of the target user’s history, resulting in personalized item representations. This study replaces the deterministic attention scores with stochastic attention weights treated as random variables whose distributions are modeled using a Bayesian approach. Through this design, the proposed model effectively captures the uncertainty stemming from implicit feedback within the vector representations of users and items. The experimental results demonstrate that the proposed model not only effectively mitigates the ambiguity of preference signals inherent in implicit feedback data but also achieves better performance improvements than baseline models, particularly on datasets characterized by high user–item interaction sparsity. The proposed model, when integrated with an attention module, generally outperformed other MLP-based models in terms of NDCG@10. Moreover, incorporating the Bayesian attention mechanism yielded an additional performance gain of up to 0.0531 compared to the model using a standard attention module.