Generalized temporal similarity-based nonnegative tensor decomposition for modeling transition matrix of dynamic collaborative filtering

In the real world, user preferences change dynamically. Therefore, time-aware recommendation systems have attracted more attention in both academia and industry. In the literature, tensor decomposition-based models and matrix factorization-based models can handle large-scale sparse data well. However, to the best of our knowledge, there is no work that provides an explanation of the latent time factor embedded in the models. Moreover, conventional Frobenius norm-based models cannot well describe the dynamic changes in user preferences over time. To capture the dynamic changes in user preferences, we interpret the time latent factor vector as a transition matrix of user preferences. In addition, a novel temporal similarity measure is proposed accordingly, which considers dynamic user and item changes between two adjacent time slices. Moreover, we propose a generalized temporal similarity-based nonnegative tensor decomposition (GTS-NTD) model and provide the corresponding solution method. Experiments on three datasets suggest that our proposed method can improve recommendation performance under dynamic changes in user preferences.