Research on the Combination Model Based on DPMM and IForest

This paper proposes a combined model based on DPMM (Dirichlet Process Mixture Model) and IForest (Isolation Forest), providing an effective solution to address the challenges of cold start and customer drift in credit risk management. Focusing on the scenario of a large amount of unlabeled data in credit risk management, the approach initially employs DPMM to cluster a small number of labeled normal and default samples. By calculating the Euclidean distance between unlabeled samples and each cluster, the model obtains a default similarity score for the unlabeled samples. Subsequently, considering the imbalance in sample distribution, the IForest algorithm constructs multiple random isolation trees to identify anomalies in the dataset, calculating the anomaly score as a measure of default abnormality. Finally, a weighted combination of the default similarity and anomaly scores from the two models is used for classification of unlabeled samples. Simulation experiments on the German credit risk dataset, with IForest model and self-training model as two control groups, demonstrate that the combined model exhibits superior predictive performance, particularly in scenarios with a low proportion of default samples, effectively mitigating the challenges associated with the cold start problem.