Prediction of bank transaction fraud using TabNet—an adaptive deep learning architecture

Battle Count: 65

The paper focuses on bank transaction fraud detection rather than quantitative trading. However, it is relevant to financial risk management, operational risk in banking, and anomaly detection in transactional data. The TabNet architecture and XAI principles could be adapted for trading anomaly detection, market manipulation detection, and risk scoring. The SMOTE balancing approach and cross-validation methodology are transferable to trading signal classification tasks. The paper's emphasis on interpretability aligns with regulatory requirements in algorithmic trading.

TabNet is available via the pytorch-tabnet library, making basic implementation straightforward. The full pipeline (data loading, categorical encoding, SMOTE balancing, 3-fold cross-validation, hyperparameter tuning, multi-model benchmarking, EDA with heatmaps/radar plots, and interpretability analysis) adds moderate complexity. The paper provides specific hyperparameters and hardware specs. However, the collapse of DNN, CNN1D, LSTM, and GRU models suggests that proper tuning of these architectures is non-trivial. The interpretability layer (feature masks) requires additional analysis steps.

The paper uses a publicly available Kaggle dataset of Indian bank transactions and implements models in PyTorch. Specific hardware (Ryzen 7, RTX 3050, 16GB RAM) and hyperparameters (max epochs 100, batch size 1024, virtual batch size 128) are provided. However, the exact Kaggle dataset link is not specified, data is available only 'on request', and no GitHub repository is provided. The 3-fold cross-validation setup and SMOTE balancing are described but some preprocessing details are vague.