Graph structures have received significant attention, owing to their natural adaptability for sessions.Thus, many researchers have investigated Graph Neural Networks(GNN)-based recommending algorithms and achieved state-of-theart performances.Existing session-based recommendations based on GNN can yield relatively accurate recommendations, utilizing structural graph information.However, they neither consider repetitive submissions from users and complex transition between items nor fully utilize complex graph structural information.Consequently, they result in prediction losses.This paper proposes a GNN sequence recommendation algorithm based on the fusion of directed and undirected information with attention. The proposed algorithm combines directed and undirected structural information of session graphs into new hidden embeddings of items.By using repetitive behavioral information and attention mechanisms, the model incorporates complex transitions of items to form better session embeddings.During feature propagation, each node strikes a balance between preserving its information and absorbing its neighbors’information, improving the accuracy of recommendation predictions. The experimental results for Diginetica, Yoochoose 1/64, and Yoochoose 1/4 data sets show that compared with the best existing algorithms, that is, Session-based Recommendation with GNN(SR-GNN)and Target Attentive GNN(TAGNN), the accuracy of the algorithm can be improved by up to 4.34%. The proposed algorithm can predict the accuracy of the user’s next click better in a session.