The session-based recommendation task is about predict the user's next action based on an anonymous session. Existing approaches all make prediction by analyzing the current session, ignoring relevant transitions between each sessions the user historically made. Therefore, a recommendation algorithm based on graph neural network to capture higher order information is proposed. Firstly, learning the embedding vector of the item by aggregating information from local neighbors as well as higher order neighbors. Secondly,capturing the short-term and long-term interest with attention mechanism. Finally generate ranking score for predition. Experimental results on real data sets Yoochoose and Diginetica show that the proposed algorithm significantly improves the performance of the recommendation system compared to exisiting session-based recommendation algorithm.