A learning framework for temporal recommendation without explicit iterative optimization

Abstract Pushing accurate and prompt recommendations to diverse users is an on-going key challenge for researchers in studying recommender systems, which is also known as temporal recommendation. Thereby, modeling to capture every single user's preference dynamically and accurately is the core. In this work, we propose a hybrid model to exploit user's periodical feature for his/her behaviors over time such as purchasing on-line goods which essentially can reflect one's preference on some types of items. An outstanding merit of the proposed model is that, in contrast to traditional model-based methods, it is free of explicit iterative optimization process. Meanwhile, we introduce an improved Markov state computing method to perform temporal recommendation on three real movie rating data sets which obtains a systematic improvement in computational efficiency compared with traditional model-based approaches and moreover, it is better than neighborhood-based collaborative filtering methods and the state-of-the-art temporal recommendation models with respect to the results of recommendation accuracy. In addition, further analyses verify that the proposed learning framework has the capacity in capturing user's personalized preference from a perspective of temporal dynamics.