Enhancing Rare Event Detection via Counterfactual Generation with Exogenous Variables

In high-risk digital environments such as financial fraud detection, cybersecurity, and modern payment ecosystems (e.g., the digital renminbi), institutions face increasing pressure to effectively identify rare but high-impact suspicious activities - despite their extremely small proportion, severely imbalanced data, and limited observable evidence - in order to protect vulnerable users from fraud and other harmful activities. Traditional methods, such as oversampling and reweighting, often struggle to capture the complex dependencies among features in minority-class samples and fail to account for unobserved exogenous variables and model explainability, especially in real-world, high-complexity scenarios. To address these limitations, we propose CARE (Causal Augmentation for Rare Events), a novel approach that jointly models exogenous variables and counterfactual data augmentation to enhance detection performance and explainability. Specifically, our approach consists of: (1) causal feature mining; (2) inferring unobserved exogenous variables via a CVAE trained on observational data; and (3) constructing an SCM to generate intervention-based counterfactual samples for rare events, enabling effective expansion of the minority class. Extensive experiments on both public and real-world datasets demonstrate that CARE significantly outperforms existing methods in terms of both detection performance and explainability. By leveraging causal modeling and counterfactual reasoning, CARE provides a theoretically grounded and empirically effective solution to rare event detection under extreme class imbalance.