Existing attention-based session recommendation models integrate all higher-order neighbor information, enriching node-item representations but also leading to representation homogenization and interference from weakly related neighbors. To address this issue, a session recommendation model based on graph attention and session preference identification is proposed. In this model, when generating node-item embeddings, only μ-neighbor are selected, and a graph attention mechanism is used to aggregate the feature information of these μ-neighbor. In addition, to account for varying session preferences across users, a method for session preference classification and embedding learning is introduced, categorizing session preferences into two types: concentrated preferences and divergent preferences, for classification computation. Finally, a soft attention mechanism is applied to integrate the item embeddings learned from the session, obtaining the session embedding representation for prediction. Experiments on two real datasets, Nowplaying and Diginetica, demonstrate that the proposed model achieves superior performance compared to baseline methods.