Automobile fraud detection with machine learning on graphsKalc, Matej (Avtor) Šubelj, Lovro (Mentor) Gorenc Novak, Marija (Komentor) Automobile fraud detectionMachine learning on graphsDeep learningSemi-supervised learningEvery year, millions of car accidents happen. Some of these are staged, faked, or with exaggerated costs. To find fraudsters, insurance companies use custom rules or machine learning algorithms. Most popular machine learning models use tabular data, which is not ideal in the context of automobile fraud detection. Fraud data is best represented as a network of nodes and edges. We compare tabular methods with heterogeneous graph neural networks on the same data set. We propose two new models: a self-supervised heterogeneous graph anomaly detector (HGAD) and a supervised heterogeneous graph fraud detector (HGFD). HGAD is has the highest anomaly detection between anomaly detectors. HGFD achieves 83\% AUC score and 0.018 Brier score, while the best performing tabular model, XGBoost, has 80\% AUC and 0.019 Brier score. The precision of HGFD and XGBoost is similar, but HGFD's precision is higher as the number of claims increases. Training HGFD with inductive learning does not improve results. The optimal solution for automobile fraud detection is a hybrid model, consisting of XGBoost and HGFD, that averages predictions of both models. This model achieves 86\% AUC, 0.017 Brier score and it outperforms the current models in use, offering a 17.4\% improvement in precision over XGBoost.20242024-09-25 15:50:01Magistrsko delo/naloga162604sl