Knowledge Contrast and Collaborative Interaction Recommender System

In recent years, knowledge graphs have been introduced into recommender systems as external information, which can optimize the precision of recommendation algorithms. However, existing knowledge graph-based recommender systems face two challenges. One the one hand, they fail to capture complex interactions at different stages in the collaborative propagation process, which prevents a deep understanding of user preferences. One the other hand, the skewed distribution of items and entities hinders the collection of sufficient supervision signals, resulting in less confident representations. To address these issues, we propose a novel recommendation model, KCCI (Knowledge Contrast and Collaborative Interaction). This model employed multi-level collaborative information propagation and dual knowledge interaction layers in the interaction between users and items. We can thus effectively model implicit complex interactions at different stages to deeply understand users' collaborative preferences for items. In addition, dual views are generated from the perspectives of user-item interaction and knowledge graph representation. These views are used to construct a dual-view contrastive learning loss, which alleviates the problem of long-tail items and entities receiving inadequate supervision signals. Experiment results on the MovieLens-lM, Last-Fm,and Book-Crossing datasets demonstrate that, compared with baseline models, our proposed model performs best and overcomes the shortcomings of the other models.