An XAI-Powered Approach for Financial Fraud Detection Using Anomaly Detection and Classification Techniques

Abstract

Introduction: With the surge in online financial transactions, fraud detection has become a critical priority. Traditional rule-based systems often fail to keep up with sophisticated fraud patterns. Moreover, the lack of interpretability in modern machine learning models poses challenges in regulated environments. This project addresses these issues by designing a transparent, intelligent fraud detection system using machine learning and Explainable AI (XAI), with a focus on both performance and usability through an integrated dashboard.

Objectives: The project aims to develop a fraud detection framework that is both accurate and interpretable. It handles data imperfections through preprocessing, detects anomalies using Isolation Forest, and confirms fraud via Random Forest. SHAP is used for model explainability, while Streamlit enables real-time interaction for end users.

Methods: The pipeline consists of several components. Data preprocessing addresses missing values using SimpleImputer, standardizes feature distributions with StandardScaler, and encodes labels using LabelEncoder. The anomaly detection layer uses Isolation Forest, which isolates rare patterns based on recursive partitioning. Flagged transactions are then passed to a Random Forest classifier for final fraud classification. For transparency, SHAP is used to explain feature contributions at both global and individual levels. These insights, along with prediction outputs, are made accessible through a Streamlit interface designed for analysts and decision-makers.

Results: The system performs effectively across all stages—preprocessing, detection, classification, and explanation—achieving high precision and recall on a highly imbalanced dataset. SHAP insights enhance model transparency, while the dashboard enables users to explore predictions and explanations in real time.

Conclusions: This modular and interpretable solution addresses both technical and regulatory requirements for financial fraud detection. Future work may focus on real-time streaming, behavioral data integration, and adaptive learning to improve performance in dynamic environments.