 |
International Society of Science and Applied Technologies |
| TreeSHAP-Guided Feature Selection for Credit Card Fraud Detection | ||||
| Author |
Qianxin Liang
|
|||
| Co-Author(s) |
Chelsea M. Zuvieta; Taghi Khoshgoftaar
|
|||
| Abstract | Credit card fraud is a persistent issue in the digital
economy, resulting in significant losses for both consumers and
financial institutions. The high dimensionality of transaction
data and the severe imbalance between legitimate and fraudulent
transactions create a complex classification problem. To
address these challenges, this study proposes an interpretable
and efficiency-driven feature selection framework based on Tree- SHAP, which derives SHapley Additive exPlanations (SHAP) values for tree-based models. Using the widely studied Kaggle Credit Card Fraud Detection dataset, the proposed framework generates TreeSHAP-guided feature rankings from three individual treebased learners: Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGB). An ensemble-based strategy is also incorporated to aggregate feature importance across the models into a single ranking. We train an RF classifier with the selected feature subsets and evaluate performance using the Area Under the Precision-Recall Curve (AUPRC) and statistical significance tests. Results show that the proposed TreeSHAP-based framework can significantly reduce the feature space while preserving strong fraud detection performance. While the ensemble-based approach provides a stable and competitive strategy, it does not achieve the best overall performance. Instead, the DT-based TreeSHAP ranking demonstrates strong transferability to the RF classifier, achieving near full-feature performance with a substantially reduced feature set. Overall, these findings indicate that a simple learner can identify highly informative features through TreeSHAP, enabling robust and efficient fraud detection models with improved interpretability.
|
|||
| Keywords | SHAP, TreeSHAP, Feature Selection, Credit Card Fraud Detection, Class Imbalance, Machine Learning | |||
| Article #: RQD2026-26 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |