Contrastive Learning based CrossDomain Recommendation via User Convergence

Cold-start users have always been a challenging task to deal with in the paradigm of recommender systems. Providentially, the presence of these cold-start users in multiple domains has addressed the problem of their sparse presence in the target domain.The interactions of cold-start users in the source domain plays a pivotal role in predicting their interests in target domain.The question of paramount importance remains what to transfer and the manner to achieve it. Most recent advancements in this area mitigate the gap of two domains by using tags as a bridge to transfer knowledge. User convergence aligns user preferences across different domains. We propose a novel framework that includes the metadata of user and items to devise a neighbourhood based on similarity of preferences they make. The semantic similarity is drawn using SBERT model with cosine similarity. This technique empirically investigates the advantage of fusing metadata through graph neural network(GNN) for recommendation tasks. Particularly, we have fused the metadata and interaction information jointly to model a graphical structure. This helps in learning a user’s representation through hierarchical graph attention model that also incorporatespreferences of likeminded users, their behaviour and rating patterns ¹ . This framework supplements the user-item ratings with embeddings generated from user’s and item’s metadata. The personalized preferences are further refined through contrastive learning. To bridge the semantic gap among two domains, a neural network is employed to learn a cross-domain mapping function. Our proposed algorithm seams the strength of GNNs with cross-domain paradigm to utilize the richness in metadatafor addressing sparsity. The combined advantages of GNNs, cross-domain and contrastive learning alleviated the issues of cold-start users by transferring user preferences from a source domain to a target domain ² .