DLIR: a deep learning-based initialization recommendation algorithm for trust-aware recommendation

摘要

In collaborative filtering recommendations, a good local minimum depends largely on the initialization of the latent feature vectors of users and items. However, many existing methods initialize them by random initialization or zero ones, which results in recommendation performance degradation. In addition, they also ignore the role of users’ trust and items’ information in obtaining reliable recommendations. Aiming at addressing these challenges, this paper proposes a deep learning-based initialization recommendation method for trust-aware recommendation, named DLIR. In DLIR framework, we first leverage deep learning to learn a better latent feature vector of users and items and then take it as the initial value of the latent factor model (LFM) recommendations. Next, the users’ trust information and items’ information are employed to construct users’ social trust ensemble and the items’ intrinsic feature relationship, respectively. Finally, the constructed information is integrated into LFM for trust-aware recommendation. To verify the effectiveness of the proposed DLIR, the extensive experiments conducted on two real datasets (e.g., Epinions and Last.fm) show that the proposed approach performs much better than state-of-the-art recommendation algorithms on recommendation accuracy.