A Study of Recommendation Algorithm Based on Graph Transformer and Contrastive Learning

This paper proposes a novel recommendation algorithm, GTransCL, which combines the benefits of Graph Transformer and contrastive learning to provide more accurate personalized recommendations. GTransCL digs deeper into the complex interactions between users and items through graph neural networks, and uses the attention mechanism to capture the key connectivity points in order to more accurately understand user preferences. In addition, GTransCL designs an integrated location coding strategy that incorporates node degree and PageRank algorithms to help the algorithm capture node information. Contrastive learning learns a more compact and discriminative embedding space by reducing the loss associated with positive sample pairs and enhancing the loss associated with negative sample pairs within the GTransCL algorithm. Graph data augmentation enhances the model’s generalization ability. Experimental results in two public datasets MovieLens-1M and LastFM compared with classical recommendation models such as MF, LightGCN, XSimGCL, etc. show that, GTransCL improves in two metrics, Recall and NDCG, which validates the effectiveness of the model.