Multi‐Task Sequential Recommendation Based on Gradient Segmentation

ABSTRACT In the real world, the interaction behaviors of users are not isolated. Instead, they occur continuously in a sequential manner. In the absence of project information, it is difficult for ordinary recommendation methods to predict their next interaction behavior, while sequential recommendation can capture the dynamic behavior of users to complete the recommendation task. Existing sequential recommendation methods generally capture more information about interactive sequences by adding auxiliary tasks for strengthening sequence representations. Despite their achievements, they are deficient in mining user‐user or item‐item multi‐level sequence information, and are not efficient in solving multi‐task optimization problems. To address the above issues, we propose gradient segmentation‐based multi‐task sequential recommendation (GS‐MTLSR). First, we cluster user interaction sequences, and then maximize the consistency between user sequence features and their corresponding behavior category features through contrastive learning. Second, we construct dynamic global graphs to capture global contextual information and enhance interaction item relationship information of the sequence through graph contrastive learning. For the optimization problem of multiple contrastive learning tasks, this paper adopts the gradient segmentation method to solve gradient conflicts to eliminate the negative cosine similarity between task optimization gradients. In this way, multiple contrastive learning tasks can be leveraged to the sequential recommendation (SR) model. Experimental results show that the recommendation accuracy of the GS‐MTLSR model achieves state‐of‐the‐art performance compared to other baseline models.