A Heterogeneous Ensemble Method for Handling Class Noise in Supervised Machine Learning

The goal of machine learning is to approximate an unknown input function by learning based on a set of labeled training samples. Noisy labels due to class noise in the training data can have three negative consequences: (i) the prediction accuracy may decrease, (ii) the complexity of the model may increase, and (iii) the number of training examples needed may increase. To tackle this problem, we present a new ensemble-based filtering approach for identifying and eliminating class noise. In our approach, we build the ensemble filter by employing k-means clustering and classifier calibration. By using a high agreement rate, our heterogeneous ensemble filter is able to collect most of the clean data. We report experiments on eight binary and five multiclass datasets from UCI benchmarks to demonstrate our proposed methods are highly effective in label noise filtering. Experimental results show that our proposed method led to significant performance improvement compared with the state-of-the-art baselines. A comparative analysis is conducted with respect to the two-stage ensemble filter, a reference homogeneous ensemble-based class noise filter, and mCRF, a reference multiclass label noise filter.