TUR: Utilizing Temporal Information to Make Unexpected E-Commerce Recommendations

Unexpectedness recommendations are getting more attention as a solution to the over-specialization of traditional accuracy-oriented recommender systems. However, most of the existing works make limited use of available interaction information to compute distance and neglect the fact that varying time intervals for recommendations would lead to different perceptions of unexpectedness from users. In this work, we propose a novel Temporal Unexpected Recommendation (TUR) approach to improve e-commerce recommendations’ unexpectedness. Specifically, we consider the complementarity of both implicit and explicit distances, modeling unexpectedness from the latent space (i.e., embedding vectors) and the side information (i.e., item taxonomy) respectively. Meanwhile, we import a module based on the time-aware GRU to leverage the impact of timeliness on recommendation unexpectedness. Experiments on a large-scale e-commerce dataset containing real users’ feedback show that TUR significantly outperforms the baselines in enhancing unexpectedness while maintaining a comparable accuracy level.