Multi-session transformers and multi-attribute integration of items for sequential recommendation

Modeling sequential dependencies plays a significant role in simulating the dynamic changes in users' interests, and the introduction of deep learning can address long sequence data to some extent, thereby enabling more precise capture of these changes. However, most existing models still struggle to train sequences with insufficient interaction information or overly long sequences, and they also fail to capture the genuine intentions of users reflected by the interaction behaviors. Additionally, they overlook the characteristic that items interacted with by users are not strictly ordered and are highly homogeneous within a certain period, while the items between different periods are likely to be heterogeneous. In this paper, we propose a sequential recommendation model based on Multi-session Transformers and multi-attribute integration of items (MTMISRec), which enriches the missing interaction information of sparse data by integrating items' attributes with users' historical interaction sequences and distinguishes the true intentions of users under similar interactions. Furthermore, we set a time threshold to partition items with interaction intervals within this threshold into a session, thereby capturing homogeneous relationships within each session. We employ the dual attention mechanism to perform local attention within each session and introduce the learned type weights of each session into the complete interaction sequence to perform global attention, thereby blurring the sequential relationships within sessions and integrating global relevance with local details to handle overly long sequences precisely. We conducted extensive experiments on four datasets, and the results demonstrate that MTMISRec surpasses advanced sequential models on sparse and dense datasets.