A Recommendation Algorithm Based on Knowledge Graphs and Improved KGCN Network

Integrating knowledge graphs as auxiliary information into recommendation systems enables recommendation algorithms to more effectively learn external knowledge and enhance the feature representation of users and items. In recent years, research on recommendation algorithms that incorporate knowledge graphs has focused on enriching the feature representation of items, neglecting the extraction of user-side features. This paper proposes a recommendation algorithm based on an improved KGCN network that retains the KGCN network's ability to aggregate item features. It replaces the part of the network that extracts user features with Long Short-Term Memory (LSTM) networks, integrates the user feature vectors from both networks, and finally uses the KGCN network's prediction layer to make predictions. This approach not only addresses the cold start problem of sparse user data common in traditional collaborative filtering recommendation algorithms but also captures the dynamic changes in user interests, enhancing the model's generalization ability to unseen data. Test results show that on the Movielens-20M dataset, the AUC increased by 1.9%, and the Fl score increased by 2.1 %. The code for the related work has been placed in the link: https://github.com/rgsyc/KGCN-LSM/tree/master.