Neighborhood structure enhancement and denoising method for multi-behavior recommendation.

Traditional recommender systems often assume that there is only one type of interaction between a user and an item, which does not reflect the complexity of real-life users engaging in multiple behaviors such as browsing, clicking, adding to cart, and purchasing. Recent multi-behavioral recommendation methods have demonstrated their effectiveness, while they still suffer from two limitations: (1) Unbalanced user interaction data and sparse node neighbor information pose challenges to user preference modeling. (2) Direct transfer of information from the auxiliary behavior to the target behavior introduces noise. In this paper, we propose a Neighborhood Structure Enhancement and Denoising method (NSED) to address such issues. NSED includes a neighborhood-enhanced Graph Convolutional Network (GCN) and a structural enhancement module to strengthen neighbor node representation and mitigate the long-tail problem. It performs cross-behavioral modeling by cascading structures to discover dependencies among different behaviors. Additionally, a denoising module is designed to alleviate the problem of model performance degradation due to the negative migration phenomenon. The user preferences learned under the target behavioral graph are shown to have high accuracy, whereas those constructed under the auxiliary behavioral graph are denoised using the contrastive learning method. Compared with the state-of-the-art (SOTA) baseline approach, NSED improves the average performance by 10.4% and 10.67% on the three public datasets. For further verification, it can be found our code and weights at https://github.com/spider-123456/NSED.