CDifOD: mitigating class imbalance in graph outlier detection through context- and category-aware diffusion

Graph-based outlier detection has gained significant attention due to its applications in security, finance, and social networks. However, the highly imbalanced nature of outlier datasets poses a fundamental challenge, leading to biased learning and suboptimal generalization. Existing strategies, including sampling-based techniques and loss reweighting, often fail to capture the intricate distribution of outliers in graph structures. To address this, we propose CDifOD, a context- and category-aware diffusion-based augmentation framework that generates realistic synthetic outliers to enhance model training. Specifically, we first apply graph clustering to partition large graphs into subgraphs, improving computational efficiency while preserving structural dependencies. Next, a GraphMAE-based encoder-decoder constructs a robust latent space for high-quality outlier generation. Finally, we introduce a context and category-aware conditioning strategy, where masked graph structures provide structural priors, and outlier labels guide the generative process to ensure realistic outlier synthesis. By integrating these components, CDifOD produces more balanced training data, mitigating class imbalance effectively. Extensive experiments on multiple benchmark datasets demonstrate that CDifOD consistently outperforms existing methods, yielding significant improvements in graph outlier detection.