Abstract

Machine learning has emerged as a powerful and indispensable tool for fraud detection due to its ability to learn complex patterns, adapt to evolving threats and make accurate prediction in real time by leveraging data-driven approaches.  Although AdaBoost has proven effective for handling imbalanced fraud datasets, it remains sensitive to noisy data and prone to overfitting. Metaheuristic approaches such as the Starfish Optimization Algorithm (SFOA) have been applied to optimize AdaBoost but suffered from premature convergence. Consequently, this study developed an Enhanced Starfish Optimization Algorithm (ESFOA) with an Elite Preservation Strategy to improve the robustness and performance of the AdaBoost ensemble model for financial fraud detection.

A financial   dataset of  ten thousand (10,000) obtained from Kaggle’s credit card fraud detection which comprises both legitimate and fraudulent transactions was utilized for the study. Preprocessing procedures included handling missing values through imputation, detecting and treating outliers using z-score and Interquartile Range (IQR) methods, performing feature selection and engineering, encoding categorical variables and scaling numerical attributes to ensure balanced and standardized input data. A baseline AdaBoost ensemble model with decision stumps as weak learners was developed, after which the Standard Starfish Optimization Algorithm was enhanced with an Elite Preservation Strategy to optimize key AdaBoost hyperparameters which include the number of estimators, learning rate, and tree depth through a defined fitness function. The enhanced algorithm iteratively evolves candidate solutions through exploration and exploitation mechanisms until convergence was achieved and the optimal parameters were used to retrain the classifier to produce the ESFOA-AdaBoost model, which was further validated using cross-validation techniques for robustness. Finally, the developed model was implemented in MATLAB (R2023a) Software and its performance was evaluated and compared with SFOA-Adaboost and Adaboost using false positive rate, sensitivity, specificity, precision, F1-score, accuracy and detection time.

The false positive rate, sensitivity, specificity, precision, F1-score, accuracy, and detection time for ESFOA-AdaBoost model were 2.22%, 98.90%, 97.78%, 99.05%, 98.81%, 98.57%, and 16.53 seconds, respectively. The corresponding SFOA-AdaBoost model and standard AdaBoost ensemble model were 6.67 and 8.67%, 97.00 and 96.14%, 93.33 and 91.33%, 97.14 and 96.28%, 96.52 and 95.48%, 95.90 and 94.70% and 24.50 and 31.32 seconds, respectively.

This developed ESFOA-AdaBoost ensemble model demonstrated superior effectiveness in financial fraud detection compared to both the conventional SFOA-AdaBoost and AdaBoost models. The developed ESFOA-AdaBoost model can serve as a reliable, scalable solution for real-time financial fraud detection systems.