A real-valued label noise cleaning method based on ensemble iterative filtering with noise score

Real-world data always contain noise for a variety of reasons. In a regression task, noisy labels interfere with the construction of an accurate model, leading to a decline in the prediction accuracy. Methods that have emerged to deal with continuous label noise are rather limited in contrast with those on class noise cleaning techniques. To address this gap, we propose a novel noise filter to clean noisy instances with real-valued label noise. This method combines several filtering strategies. First, an iterative filtering process is carried out, allowing us to avoid using potential noisy examples in each new filtering iteration. Second, we develop a noise score to assess the noise level of each detected noisy instance. The higher the noise score is, the more likely that the instance is noisy. Finally, an ensemble filtering scheme is implemented. The fusion of detection from different models makes the determination of noisy examples even more reliable. The validity of the proposed method is verified through extensive experiments. We discuss the selection of the best hyperparameters, and compare the developed method with several state-of-the-art noise filters using public regression datasets. The outcomes show that our method not only achieves a good balance between the elimination of noisy samples and the retention of clean samples but also outperforms all the other compared methods, especially at higher noise levels. Simultaneously, the results of a case study of temperature prediction in an electric arc furnace suggest that training a domain-related regressor on a dataset preprocessed with the proposed noise filter contributes to a great improvement in the prediction accuracy.