This paper proposes a novel session-based recommendation method called SR-GNN, which uses graph neural networks (GNNs) to model session sequences as graph-structured data. Traditional methods model sessions as sequences and estimate user and item representations, but they struggle to capture complex item transitions and accurately estimate user vectors. SR-GNN models sessions as graphs, allowing GNNs to capture complex item transitions that are difficult for conventional sequential methods. Each session is then represented as a combination of global preference and current interest using an attention network. Extensive experiments on two real-world datasets show that SR-GNN outperforms state-of-the-art session-based recommendation methods consistently. The method models session sequences as graphs, captures complex item transitions, and uses attention mechanisms to combine global and local session preferences. It also uses soft-attention to generate session embeddings that automatically select significant item transitions and ignore noisy user actions. The proposed method is flexible in constructing connection schemes between items in the graph and performs well across different session lengths and embedding strategies. The results show that SR-GNN achieves better performance than other methods in terms of precision and mean reciprocal rank. The method is effective in capturing complex user behavior and provides more accurate recommendations.This paper proposes a novel session-based recommendation method called SR-GNN, which uses graph neural networks (GNNs) to model session sequences as graph-structured data. Traditional methods model sessions as sequences and estimate user and item representations, but they struggle to capture complex item transitions and accurately estimate user vectors. SR-GNN models sessions as graphs, allowing GNNs to capture complex item transitions that are difficult for conventional sequential methods. Each session is then represented as a combination of global preference and current interest using an attention network. Extensive experiments on two real-world datasets show that SR-GNN outperforms state-of-the-art session-based recommendation methods consistently. The method models session sequences as graphs, captures complex item transitions, and uses attention mechanisms to combine global and local session preferences. It also uses soft-attention to generate session embeddings that automatically select significant item transitions and ignore noisy user actions. The proposed method is flexible in constructing connection schemes between items in the graph and performs well across different session lengths and embedding strategies. The results show that SR-GNN achieves better performance than other methods in terms of precision and mean reciprocal rank. The method is effective in capturing complex user behavior and provides more accurate recommendations.