Fraud Detection with Machine Learning in Property Insurance Policy Requests
Sergül Ürgenç, H. Eliot Kaplan, Ayça Çakmak Pehlivanlı · Zenodo (CERN European Organization for Nuclear Research) · 2022
The purpose of this study is to predict in advance, whether the policy claims are made for abuse in order to reduce claim payments in the property branch of the insurance industry and to prevent any financial losses. In order to detect abnormal situations, Label Spreading and Self Training semisupervised machine learning approaches were used due to the insufficient label ratio of the anomaly class in the data set. After the data labeling process, fraud prediction study was conducted with supervised machine learning models and it was discussed which semi-supervised learning approach worked with higher performance. The datasets include the policy demands in 2017 and 2018 and the weather information of that location. Accuracy, precision, recall, specificity, and F1 score were evaluated as model success measures, and according to these results, it was seen that the Self Training approach could label data with higher performance than the Label Spreading approach.