A Knowledge Graph-Enhanced Recommender System Based on Variational Graph Autoencoder

Recommender systems have gained widespread adoption but still suffer from challenges such as limited feature diversity and difficulties in capturing complex user–item interactions. In this article, we propose KG-EVGAE, a knowledge graph (KG)-enhanced recommender system based on a variational graph autoencoder. KG-EVGAE first applies a path augmentation strategy to enrich the KG with potentially missing relations, thereby enhancing entity connectivity and revealing latent structural semantics. Based on the augmented KG, convolutional neural networks are used to extract interaction features, while an attention mechanism is employed on the user social network to adaptively learn user features. These features are fused through multiple feature fusion strategies and fed into an improved variational graph autoencoder to learn expressive latent embeddings for recommendation. Thus, KG-EVGAE not only mitigates the challenges of missing features and diversity insufficiency but also captures complex interactions and dependencies between nodes and edges more effectively. Extensive experiments con ducted on four standard datasets demonstrate the superiority of KG-EVGAE in recommendation tasks.