Abstract:
Predicting users’ next-click according to their historical session records, also known as session-based recommendation, is an important and challenging task and has led to a considerable amount of work towards this aim. Several significant progresses have been made in this area, but some fundamental problems still remain open, such as the trade-off between users’ satisfaction and predictive accuracy of the models. In this study, we consider the problem of how to alleviate user interests drift without sacrificing the predictive accuracy. For this purpose, we first set up an item dependency graph to represent the click sequence of items from a global, statistical perspective. Then an efficient graph embedding learning algorithm is proposed to produce item embeddings which preserve the information flow properties of the system and the structural dependency between each pair of items. Finally, the proposed model is capable of capturing the users’ general interests and their temporal browsing interests simultaneously by using of a BiLSTM based long/short term memory mechanism. Experimental results on two real-world data sets show that the proposed model not only performs better in terms of predictive accuracy but also demonstrates better diversity and novelty in its recommendations as compared with other state-of-the-art methods.