IDOGHO JOSHUA OSHOZE

The exponential rise in digital banking transactions has heightened the risk of fraudulent and anomalous bank transfers, necessitating intelligent and automated mechanisms for its detection. Traditional rule-based systems often fail to capture evolving fraud patterns, motivating the adoption of machine learning-based anomaly detection techniques. This study aims to develop and evaluate robust unsupervised learning models for detecting anomalies in bank transaction datasets. Specifically, the research applies three state-of-the-art algorithms; Isolation Forest, Local Outlier Factor (LOF), and One-Class Support Vector Machine (OCSVM) to identify irregular transaction behaviors indicative of potential fraud. The methodology involves comprehensive data preprocessing, scaling, encoding, and feature selection to improve model learning. Real-world bank transfer datasets from kaggle were utilized, for this training. Each model’s performance was assessed using standard evaluation metrics, including; Silhouette Score, Anomaly Ratio, and Average Decision Score. Results show that OCSVM performed best (Silhouette = 0.635, strong decision scores), reliably flagging about 5% of records as anomalies. Analysis of flagged transactions revealed consistent patterns high transaction amounts or balance-change ratios, very short or very long intervals between transactions, and activity at unusual hours making the alerts interpretable and practically useful.