Multi-task intent recommendation based on dynamic and static intent integration and disentanglement

Recommender systems that mine users’ intentions to explore their potential interaction preferences have received increasing attention. However, the existing research on intent recommendation has some limitations. On one hand, the extant studies only consider users’ historical interaction information and the sparse interactions cannot reflect users’ potential interaction intention; on the other hand, they do not consider the changes in users’ kinematic and static intentions over time and the importance of users’ intention disentanglement representation, which makes it impossible for the general intention recommendation model to obtain a better representation of the intention. We propose a multitask recommendation model with dynamic and static intent integration and de-entanglement. The model mines users’ dynamic and static intents and then combines them with regularization to model the independence of the intents, encouraging the differences between the intents. Meanwhile, to further alleviate the data sparsity problem, this study additionally constructs user–user and item–item graphs using four different similarity measures, such as cosine similarity and mutual information, applies graph convolutional networks to learn about the three graphs, and then captures the complementarity between different graphs using a graph-level cross-attention mechanism. Extensive comparative experiments and ablation studies on three public datasets demonstrate that DSI-ID consistently outperforms all baseline methods, achieving a 3.5%–7.4% improvement in recommendation performance over the best baseline.