SA-Diff: Semantic-Aware graph outlier generation via diffusion models for graph out-of-Distribution detection

Graph learning has demonstrated remarkable success in modeling relational data, yet its performance often degrades when encountering out-of-distribution (OOD) samples—a critical challenge in real-world applications like drug discovery and disease diagnosis. Existing OOD detection methods for graph-structured data lack explicit guidance from OOD signals with semantic shifts, resulting in suboptimal decision boundary and limiting adaptability in dynamic environments. To address above limitations, we propose the SA-Diff framework. By guiding the generation process of graph diffusion models with semantic shifts, SA-Diff synthesizes OOD samples that exhibit diverse structural and semantic deviations. These synthetic outliers are then leveraged for boundary-aware fine-tuning, encouraging tighter and more discriminative separation between in-distribution (ID) and OOD classes. Moreover, we introduce a dual-driven scoring function that jointly considers the characteristics of ID data and generated OOD samples, leading to enhanced detection performance. Extensive experiments on molecular and biological graph datasets demonstrate that SA-Diff significantly outperforms state-of-the-art methods in OOD detection performance.