The sequential pattern behind users’ behaviors indicates the importance of exploring the transition relationships among adjacent items in next-item recommendation task. Most existing methods based on Markov Chains or deep learning architecture have demonstrated their superiority in sequential recommendation scenario, but they have not been well-studied at a range of problems: First, the influence strength of items that the user just access might be different since not all items are equally important for modeling user’s preferences. Second, the user might assign various interests to certain parts of items, as what often attracts users is a specific feature or aspect of an item. Third, many methods ignore the complex item relations in user’s previous actions. In this paper, we present a novel recommendation approach with gating mechanism and encoding module to address above problems. Specifically, the pair-wise encoding layer is first introduced to build 3-way tensor for modeling the relationships among items in user interact histories. We also apply two gating layers to filter useful information and capture user’s short-term preference from aspect-level and item-level. We also follow the similar sprits to model user’s long-term preference by integrating user latent embeddings. Empirical results on three public datasets show that our method achieves effective improvements over the state-of-the-art sequence-based models.