Enhancing recommendations with contrastive learning from collaborative knowledge graph

There have been excellent results using knowledge graphs in recommender systems. Knowledge graphs can be used as auxiliary information to alleviate data sparsity and strengthen the modeling of item sets and the representation of user preferences. However, users as the Core subject in the recommendation process, should be taken seriously. We believe that the user's choice of items will be affected by internal and external factors. Internal factors refer to the users’ fuzzy interest sets, which initially affect the users' choices. External factors refer to the influence of similar users and similar items in the users' selection of items. Inspired by the success of contrastive learning in graph collaborative filtering, we propose the Knowledge Augmented User Representation (KAUR) model to explore contrastive learning in collaborative knowledge graphs, learning semantic neighbors (external factors) and extract fuzzy interest sets (internal factors) from collaborative knowledge graphs. Specifically, we use the graph neural network to learn the representation of each node in the collaborative knowledge graph and regard the information of nodes and their propagated neighbors’ information as positive contrastive pairs, and then use contrastive learning to enhance the node representations. To further explore the potential interests of users, we regard users (or items) with other similar users (or items) as semantic neighbors and incorporate them into contrastive learning as positive pairings as well. Then the extracted fuzzy interest sets are merged into the user representations to get better interpretability. We conduct extensive experiments on three standard datasets and the results show that our KAUR model outperforms current state-of-the-art baselines.