Development of Fraud Detection Model Using Support Vector Machine and Constraint Solver

Abstract

The expansion of digital payments has led to a rise in sophisticated fraud, creating major financial and reputational risks for institutions. Although Support Vector Machines (SVMs) have demonstrated strong predictive performance in fraud detection, their decision-making processes often lack transparency and may not fully align with domain knowledge or regulatory requirements. In contrast, constraint solvers are inherently explainable and exhibit strong compliance with predefined rules, but they lack adaptability to complex, data-driven patterns. This paper addresses this limitation by developing a hybrid fraud detection model that integrates Support Vector Machines with Constraint Solvers so as to enhance fraud detection performance. SVM model was used to identify patterns distinguishing fraudulent from legitimate transactions based on historical data. Simultaneously, a constraint solver model was designed using rules derived from transaction characteristics, principal component analysis (PCA) value ranges, and anomaly detection results from an Isolation Forest. The Kaggle Credit Card Fraud Detection dataset was used for experimentation. Hybridization was achieved by combining the outputs of the SVM and constraint solver using logical OR and AND operators, resulting in two hybrid models. The implementation was done using Google Colab. Model performance was evaluated using precision, recall, F1- score, and confusion matrix metrics. The experimental results indicate that SVM model achieved strong performance, with a precision of 60.29%, recall of 83.67%, and an F1-score of 70%. The constraint-only model recorded a precision of 4.26%, recall of 63.27%, and an F1-score of 7.98%. The Hybrid (ML OR  Constraints)model achieved a precision of 5.52%, recall of 83.67%, and an F1-score of 10.36%, while the Hybrid (ML AND Constraints) model attained a precision of 57.94%, recall of 63.27%, and an F1-score of 60.49%. In summary, according to the result, SVM model performed better than hybrid model in fraud detection.