Self-filtering Residual Attention Network Based on Multipair Information Fusion for Session-Based Recommendations

Neural network session-based recommendation (SBR) models are challenged to capture transformation relationships in a chain of anonymous user activities (i.e., interaction) to predict the next interact item in the session. However, under the auspices of user anonymity and short activity durations, data sparsity is a significant problem for these models. Moreover, given that human users rarely follow a scripted session, many noisy interact items can exist in a given series, which propagates irrelevant information to the models’ dissemination and aggregation stages, leading to poor predictions. Furthermore, correlations between interaction cannot be assumed to be limited to adjacent actions. To mitigate these tough SBR challenges, this study provides a novel self-filtering residual attention network based on multipair information fusion—the MIF2-RAN. Within the model, a multipair information fusion module applies a gated recurrent unit variant to fuse multipair information and enhance session item representations. A self-filtering residual attention module then filters noisy data and learns the global dependencies of the valid interactions. Experiments on three real-world datasets show that the performance of our MIF2-RAN surpasses that of many current state-of-the-art models.