A Two-Stage Multi-Class Classification Approach Based on Anomaly Detection

In this study, we propose a new two-stage machine learning based training algorithm that aims increasing multiclass classification performance through incorporating anomaly detection. In the first stage, separate anomaly detectors are trained for each class. Then, a multi-class classifier is trained using the obtained anomaly scores obtained during the first stage in addition to the raw features. The performance of the proposed approach is analyzed over numerous datasets and compared with traditional methods. The proposed model significantly improves the performance of traditional multi-class classifiers.