Optimization of Logistic Regression Algorithm Using Grey Wolf Optimizer for Credit Card Fraud Detection

Abstract

Purpose: The advancement of digital technology has significantly changed the financial transaction system, but has also led to an increase in cybercrime, especially credit card fraud. This crime poses a significant financial threat, with reported losses reaching hundreds of millions of dollars annually. This study aims to improve the effectiveness of fraud detection using the Logistic Regression (LR) algorithm, which although widely used in binary classification, is still vulnerable to challenges with imbalanced data. The goal is to optimize LR using the Grey Wolf Optimizer (GWO) to improve accuracy and reliability.

Methods: This research implements a Logistic Regression (LR) model whose hyperparameters are optimized using Grey Wolf Optimizer (GWO) algorithm. The model was trained and tested on a public Kaggle dataset containing 284,807 credit card transactions. Data preprocessing includes handling outliers using Interquartile Range (IQR) method and handling class imbalance using KMeansSMOTE. Evaluation metrics include accuracy, precision, recall, f1-score, and specificity based on confusion matrix.

Result: The baseline LR model achieved 99.92% accuracy, 75.18% precision, 74.73% recall, 75.45% F1-score, and 99.96% specificity. After GWO optimization, the model improved to 99.94% accuracy, 85.96% precision, 83.08% recall, 84.01% F1-score, and 99.97% specificity, showing a significant performance boost. This represents a notable improvement in key metrics for fraud detection, with an increase of 14.3% in precision, 11.2% in recall, and 11.3% in the F1-score, demonstrating a more robust model.

Novelty: This study proposed the application of the Grey Wolf Optimizer (GWO) for hyperparameter tuning of a Logistic Regression model in the context of fraud detection. Unlike conventional optimization techniques that can be computationally expensive, our GWO-based approach offers an efficient and effective method for discovering optimal model settings. The optimized model not only outperforms the baseline LR but also presents a scalable and powerful solution for financial institutions to improve the accuracy of their fraud detection systems.