Learnable Imputation and Bilinear Cross-View Encoding for New-Account Fraud Detection

New account fraud poses a persistent challenge in modern banking systems due to the sparsity, heterogeneity, and incompleteness of user information. Existing methods often struggle with missing data, limited cross-view representation, and weak adaptability to evolving fraud patterns. To address these issues, we propose NAFNet, a deep learning framework that integrates dynamic feature imputation, multi-view encoding, and attention-based representation learning. NAFNet employs a learnable imputation module guided by statistical priors, encodes heterogeneous views via dedicated encoders with bilinear fusion, and enhances global dependency modeling through attention-augmented neural layers. A fine-tuned training regime ensures robustness and generalization. Experiments show that NAFNet offers substantial improvements over conventional methods, demonstrating its effectiveness in complex, real-world fraud detection scenarios.