Financial Risk Identification Using Unified Multi-Source Feature Learning and Structural Aggregation

Abstract

This paper proposes a supervised risk probability modeling framework based on multi-source feature fusion for financial risk identification tasks, enabling stable identification in business environments characterized by high noise, strong nonlinearity, and subject heterogeneity. The method takes multi-source data, such as transaction and identity data, as input, obtaining consistent numerical representations through a unified standardization and missing data correction process. A lightweight nonlinear encoder is then employed to learn risk-related latent space features. When structural information is available, neighborhood aggregation is introduced to inject local contextual relationships, thereby simultaneously characterizing individual attributes and interactive influences. Subsequently, the model maps the fused representation to risk logarithms and outputs risk probabilities via a Sigmoid algorithm, facilitating integration with threshold-based early warning and risk stratification strategies. During the training phase, weighted binary cross-entropy combined with a regularization term is used to adapt to class imbalance and differences in misjudgment costs, improving the robustness of the discrimination boundary. Comparative experiments on public datasets demonstrate that the proposed method outperforms existing representative models in accuracy, precision, recall, and overall metrics, validating the effectiveness and versatility of the framework in multi-source fusion and risk identification.