Firefly Optimization-Based Logistic Regression Classifier for Credit Card Fraud Detection
Bharti Chugh, Nitin Malik, Deepak Kumar Gupta · Journal of Circuits Systems and Computers · 2025
By detecting and preventing fraud, individuals can avoid incurring unauthorized charges on their accounts, whereas financial institutions can reduce their losses and maintain consumer trust. It is necessary to examine transaction patterns to identify purchases that exhibit variations from the cardholder’s typical purchase patterns, are a disproportionately large percentage of the cardholder’s typical purchases, are made at odd hours, are made multiple times in a short period, or exceed a certain threshold. Since it can swiftly analyze enormous volumes of data and spot trends and conclude that data, machine learning is highly suited for spotting fraudulent transactions. Within the scope of this work, a complete method for identifying fraudulent credit card activity is proposed. Yeo–Johnson transformation is applied to a standardized dataset. Ten optimization algorithms are applied for optimal feature selection. Working with optimized selected features rather than all features reduces the danger of overfitting and improves model performance. The bias in the dataset is mitigated via a reweighing algorithm. Stratified 10-fold cross-validation is applied to reduce overfitting. Fourteen classifiers are generated on an unbalanced dataset, and 210 classifiers are generated on the balanced dataset by applying 15 resampling techniques on each of those 14 classifiers. The computational results demonstrate that the suggested SelectKBest-BorderlineSMOTE logistic regression classifier with a Mathew correlation coefficient value of 0.4449 outperforms all other classifiers and is superior to 11 previous works in the literature. The Wilcoxon rank test is used to assess the model’s success statistically. SHAP AI is used to show how well the model works. Our proposed method, which integrates firefly optimization with logistic regression, yields a significant performance improvement, achieving a 34.94% increase in the MCC compared with traditional feature selection techniques.