BMDF-SR: bidirectional multi-sequence decoupling fusion method for sequential recommendation

In the domain of sequence recommendation, contextual information has been shown to effectively improve the accuracy of predicting the user’s next interaction. However, existing studies do not consider the dependencies between contextual information and item sequences, but the contextual information is directly fusing with the item sequences, which brings the problems described below: (1) Direct fusion fuses contextual information (e.g., time and categories) with item sequences which increases the dimensionality of the embedding matrix, thus increasing the complexity of the attention computation. (2) The attention computation of heterogeneous context information in the same embedding matrix makes it difficult for the recommendation model to distinguish this heterogeneous information. Therefore, we propose a bidirectional multi-sequence decoupling fusion method for sequence recommendation (BMDF-SR) to address the above issues. To establish the dependencies between temporal context sequences and item sequences, we first treat temporal contextual information as independent sequences and build bidirectional dependencies between contextual information sequences and item sequences via a three-layer seq2seq structure. Then, we perform attention computation independently for context sequences such as categories, and the complexity of attention computation can be effectively reduced by this decoupled attention computation. Moreover, since the attention computation is performed separately for each sequence, the interference between heterogeneous information during sequence fusion is reduced, allowing the model to effectively discriminate between different types of information. Extensive experiments on four real-world datasets show that the BMDF-SR method outperforms popular models.