Item attributes fusion based on contrastive learning for sequential recommendation

Sequential recommendation aims to recommend the next item for a user to interact with by analyzing the user’s historical interaction sequences. Recent studies have utilized attribute information to enhance the performance of recommendation systems. However, these studies can not fully consider the impact of different attributes on item representations. To alleviate this problem, we propose Item Attributes Fusion based on Contrastive Learning for Sequential Recommendation (IAFCL). Specifically, we design an attribute fusion module, which assigns specific weights to various attributes through a constructed item-attribute bipartite graph. Subsequently, it performs a weighted summation of all attributes and their item embeddings to enrich the item representation. In addition, we propose a hybrid loss function that includes contrastive loss function. In constructing the contrastive loss function, we employ a new combined data augmentation strategy to enrich the diversity of contrast samples, thus enabling the model to learn more differentiated representations. Experimental results on three public datasets show that IAFCL outperforms benchmark models.