Interpretable Model For Anomaly Detection In P2P Financial Transactions

Abstract

The rapid growth of peer-to-peer (P2P) digital transactions has increased the risk of transaction errors and anomalous activities, which presents a requirement for a robust and understandable system for detecting anomalies. Although machine learning models have shown promising results in anomaly detection, their integration in real world setting is limited due to lack of transparency. This study focuses on developing an anomaly detection framework for P2P financial transactions utilizing a set of contextual, behavioral, network, and temporal features using financial fields. Additionally, several preprocessing techniques are applied to deal with class imbalances. This study utilizes Machine Learning models such as Logistic Regression, Decision Trees, Random Forest, and XGBoost. XGBoost achieves the best overall performance with a ROC-AUC of 0.968 and a PR-AUC of 0.076, representing a 69-fold improvement over the random baseline. A recall of 0.943 indicates that 94.3% of anomalous transactions are correctly identified. Furthermore, to enhance model interpretability, Explainable AI (XAI) techniques such as SHAP, LIME, and CIU are applied to the best performing model (XGBoost). Spearman rank correlation analysis confirms strong inter-method agreement between SHAP and CIU (ρ = 0.745, p < 0.001), providing quantitative evidence of explanation consistency.