Heterogeneous semantic graph embedding assisted edge-sensitive learning for cross-domain recommendation

Abstract

In the digital age, recommendation systems navigate vast alternatives. Content-based, collaborative filtering, deep-driven, and cross-domain recommendation (CDR) have been studied significantly but face cold-start and data sparsity. Though CDR methods outperform others, they struggle to optimize user-item matrices. Recent graph-based CDR methods improve efficiency by leveraging additional user-item interactions; however, optimizing graph features remains an open research area. Moreover, current techniques do not consider the impact of noise items (unrelated) on recommendation accuracy. To address this gap, this paper develops a heterogeneous semantic graph-embedding (HSGE) edge-pruning model that leverages user ratings and item metadata in the source and target domains to recommend items to target domain users. To achieve it, at first Word2Vec method is applied to explicit and implicit details, followed by Node2Vec-driven graph embedding matrix generation. Our HSGE method obtains user-user, user-item, and item-item connections to achieve more semantic features. To improve accuracy, our model prunes edges that drop source domain items and allied edges unrelated to the target domain users. Subsequently, the retained HSGE matrices from both domains are processed for element-wise attention. A multi-layer perceptron with cosine similarity processed combined features matrices to generate top-N recommendations with superior hit-rate (HR) and normalized discounted cumulative gain (NDCG).