Financial risk scoring is a cornerstone of modern digital lending and payment systems, where automated risk scoring must identify rare but high-impact malicious behaviors among massive volumes of legitimate activity. This task is challenging due to heterogeneous tabular inputs, severe class imbalance, heavy-tailed noise, pervasive missingness, and distribution shifts that undermine model stability in deployment. Existing solutions span expert-rule and statistical models, gradient-boosted decision trees, and deep tabular architectures such as interaction networks and attention-based Transformers. However, tree ensembles often depend on manual feature engineering and may under-capture high-order cross-field mechanisms, while many deep interaction models remain sensitive to irregular value distributions and missing data. Moreover, their learned risk factors are typically implicit, complicating auditing and producing fragile explanations under operational constraints. Interpretable additive models improve transparency but struggle to recover performance when the underlying risk signal arises from complex feature interactions. To address these limitations, we propose ConvKAN, a dual-branch fraud scoring framework that couples convolution-enhanced interaction learning with explicit marginal-effect modeling via Kolmogorov–Arnold Networks (KAN). ConvKAN integrates robust training-statistics-driven preprocessing and field tokenization with a learnable risk token, stacked ConvFusion blocks combining multi-head self-attention and lightweight convolutional token mixing, risk-token-guided attention pooling, and a parallel spline-based KAN branch whose outputs are fused by a compact prediction head. Experiments on two public real-world datasets, Give Me Some Credit (GMSC) and LendingClub, demonstrate the effectiveness of ConvKAN. On GMSC, ConvKAN achieves an F1-score of 0.826, outperforming strong deep tabular baselines (0.821) and standalone KAN (0.618). Ablation studies further verify the individual contributions of ConvFusion, the KAN branch, and risk pooling to accuracy and robustness.
Fan et al. (Thu,) studied this question.