Predictive Model for ATM System Defects Using Machine Learning: A Recommendation for ATM Vendors

Recent research highlights the significant challenge posed by major ATM downtime resulting from system defect incidents, impacting users, ATM owners, banks, and service providers. The unpredictability of incidents adversely affects the global delivery of financial services, leading to dissatisfaction among clients engaging in financial transactions through ATMs. This paper proposes a Machine Learning (ML) approach, specifically utilizing Support Vector Machines (SVM), to predict and diagnose defective components at the onset, aiming to counteract the downtime experienced by ATM users, service providers, and banks. The study employs a 2018 real-world dataset of ATM system defects incidents from NCR in Ghana, implementing SVM algorithms fine-tuned through hyperparameter optimization in Python 3 on Google Colab. The research addresses the challenge of efficiently classifying ATM system defects in a multiclass setting, considering the limitations posed by existing approaches when dealing with numerous classes. The paper introduces the One-Versus-One (OVO) decomposition technique in conjunction with SVM to enhance the classification of multiclass datasets. Initial evaluation of three SVM kernel classifiers - linear (70.6%), polynomial (72.56%), and radial basis function (RBF) kernel (81.21%) - reveals the need for improved accuracy. Hyperparameter optimization is applied, resulting in enhanced classification accuracies for SVM-Linear (76%), SVM-Polynomial (77%), and SVM-Radial Basis (86%). The study demonstrates a substantial decrease in computational time alongside improved classification performance.