Session-based recommendation aims to infer user preferences from short-term anonymous interaction sequences, but data sparsity and limited context pose significant challenges. Existing methods often model sessions from a single perspective, overlooking rich behavioral patterns, trend information, and historical influences. To address these limitations, we propose a scalable \({\textbf {C}}\) ontrastive \({\textbf {U}}\) nified- \({\textbf {L}}\) evel \({\textbf {T}}\) rend- \({\textbf {A}}\) ware graph architecture for session-based recommendation (CULTA), a unified-level framework that integrates a session-level and a global-level graph to capture item transitions and cross-session dependencies, while a community-guided hypergraph captures high-order community semantics through contrastive learning. Furthermore, a dynamic datastore stores historical session and label embeddings during training, enabling trend-aware retrieval of positive and negative neighbor sessions for collaborative filtering. Due to its multi-level graph modeling, contrastive learning and datastore retrieval mechanism over large-scale session data, CULTA is well suited for high-performance large-scale recommendation settings. Experiments on three real-world datasets verify that CULTA significantly outperforms state-of-the-art baselines.
