Enhancing Credit Card Fraud Detection with Clickstream-Based Behavioral Features

In 2022, global losses from credit card fraud reached $33.45 billion, highlighting the need for advanced detection systems. Recent studies suggest that customer clickstream data - sequences of online interactions - can improve fraud mitigation by incorporating behavioral patterns alongside transactional information. This research evaluates the performance of XGBoost, CatBoost, LSTM, and Random Forest in detecting fraud using a transactional dataset in 2022, enhanced with user clickstream variables. Two scenarios were designed: the first using only transactional data, and the second combining both transactional and clickstream data. Preprocessing included SMOTE-ENN for class balancing, feature engineering, standardization, and One-Hot Encoding (except for CatBoost). Models were trained using K-Fold cross-validation and optimized via Bayesian tuning, with a 70-30 train-test split. Results show that XGBoost performed best in Scenario 1, with 94% accuracy and a 64% F1-Score for fraud. In Scenario 2, CatBoost achieved 96% accuracy and a 78% F1-Score, outperforming all models. These findings demonstrate that integrating clickstream behavior significantly improves fraud detection, offering a robust tool for mitigating financial risk in the banking industry.