A systematic review of the literature on deep learning approaches for cross-domain recommender systems

Abstract

The increase in online information and the expanding diversity of user preferences require developing improved recommender systems. Cross-domain recommender systems (CDRS) have emerged as a favorable solution to solve issues related to cold start, data sparsity, and diversity by leveraging knowledge from the source domains. This systematic literature review delves into the latest deep learning approaches utilized for CDRS, comprehensively analyzing state-of-the-art techniques, methodologies, metrics, datasets, and applications. We systematically review selected primary studies from popular databases covering sixty-eight publications from 2019 to March 2024. The review process involved selecting relevant studies based on the predefined inclusion and exclusion criteria to ensure the inclusion of high-quality research. Key deep learning (DL) models explored include neural collaborative filtering, convolutional neural networks, recurrent neural networks, variational autoencoder, and generative adversarial networks. We also examine the hybrid models that integrate DL with traditional machine learning techniques to enhance recommendation performance. Our findings reveal that DL approaches significantly improve accuracy, cold start, and data sparsity. This review also identifies current trends and future research directions, emphasizing the potential of Artificial Intelligence (AI), transfer learning, and reinforcement learning in advancing CDRS. In our analysis, we discovered that the domains mainly utilized are movies, books, and music, respectively, and the most widely used evaluation metrics are root mean square error (RMSE) and normalized discounted cumulative gain (NDCG). Research challenges and future scope are also highlighted to assist the researchers and practitioners seeking to develop robust cross-domain recommender systems using DL techniques.