Abstract: Session-based recommendation aims to predict users’ actions based on anonymous sessions. The existing work lacks the dynamics of user interest, user intention and the unity of capturing local and global dependencies between projects. To solve this problem, a new target aware self-attention network for session-based recommendation is proposed. It captured the global dependencies among items by combining self-attention networks, and captured the complex transitions patterns among items by using graph neural networks. The dynamic interest module was introduced to model the change of user’s interest with respect to varied target items, to model the dynamics of user’s interest, and to capture the intention characteristics behind user’s behavior. Dynamic interest representation and user feature representation were combined to obtain dynamic session representation for subsequent prediction. Experimental results on two real datasets show that the proposed model is better than the baseline method.