An embedded boosting approach incorporating instance valuation and adaptive cleaning for imbalanced credit classification with label noise
Аннотация
Class imbalance is a crucial problem in credit risk assessment, where model predictive accuracy is compromised not only by imbalanced data but also by challenges related to data traits (e.g., noise). Class imbalance combined with label noise aliasing can cause the model to overfit noisy instances, leading to reduced classification performance and potential economic losses. To address these issues, an embedded boosting approach incorporating instance valuation and adaptive cleaning is proposed. In this methodology, the imbalanced classification entropy (ICE) index is first designed in the Boosting with Instance Valuation (Boosting-IV) algorithm, which is used to calculate the base classifier weights of each layer iteration in the hybrid ensemble classifier (HEClassifier). Accordingly, the instance valuation and the corresponding weight of the HEClassifier on the final ensemble classifier are calculated. Subsequently, the Boosting with Adaptive Cleaning (Boosting-AC) algorithm identifies and processes noise based on the instance evaluation from the Boosting-IV stage, enhancing instance information and improving model classification performance. Finally, the ensemble classifier is constructed by combining the trained HEClassifier for testing. To verify the effectiveness of the proposed embedded boosting strategy, some experiments are conducted on three imbalanced credit datasets. The experimental results show that the proposed embedded boosting model outperforms the other benchmark models listed in this study, indicating its effectiveness as a credit risk classification tool in the presence of class imbalance and label noise aliasing.