I. Introduction
In the era of information overload, the personalized recommendation has been an essential component in various commercial applications, which could push personalized information for customers and increase great profits for content providers. Learning users’ preferences toward items from historical interaction records is the core of a personalized recommendation. Traditional recommender systems usually model and understand user behaviors in a static manner, ignoring the fact that user interests generally shift over time. In recent years, due to the high practicability, sequential recommendation greatly attracts academia and industry concerns. Many efforts toward this task have been devoted to developing various sequential models, such as Markov chains (MCs) techniques [1], [2], recurrent neural networks (RNNs) [3], [4], convolution neural networks (CNNs) [5], [6], and self-attention mechanisms [7], [8]. Nonetheless, the existing sequential recommenders mainly capture item transition patterns within individual sequence, which is not expressive enough to model short-term user interest especially for cold-start users.