Real time fraud detection using machine learning

Fraud detection is one of the most critical issues in the insurance industry. While current literature often presents theoretically valid models, only a few have utilized real-world data and others frequently overlooked the actual balance of target variable. Additionally, the lack of transparency and interpretability in machine learning models predicting insurance fraud stands out as another area that needs improvement. The main challenges of fraud detection include dealing with highly imbalanced datasets and interpreting model metrics properly. Emphasizing the explainability of models, both in the data preparation process and the evaluation of results, contributes significantly to the consistency of decisions, accountability requirements, and the enhancement of business knowledge. This thesis aims to provide a methodology for creating explainable machine learning models using real-world insurance data (with a target variable rate of 0.5 percent). Effects of various feature selection methods on performance have been tested with various machine learning algorithms, and the most suitable machine learning algorithm-feature selection method pair has been attempted to be identified. Among the ML algorithms that yielded the best results, a meta-ensemble self-learning model approach was proposed using a voting mechanism; the results were optimized with model threshold value. In evaluating the model results, it should be considered that a highly imbalanced dataset is being used. In light of this criterion, the precision, sensitivity, and F1-score values of the baseline model created were calculated as 28.02%, 38.41%, and 32.4%, respectively. The results obtained using the proposed meta-ensemble self-learning method were increased by 22.7%, 11.3%, and 31.8%, respectively. The interpretability of the model obtained with the baseline model approach was evaluated by examining SHAP values, analyzing the similarities and divergences of the features used in the model. The tests indicate that the model obtained with the proposed method improves fraud prediction performance without compromising consistency.