Similarity-based and Iterative Label Noise Filters for Monotonic Classification

Abstract

Monotonic ordinal classification has received an increasing interest in the latest years. Building monotone models from these problems usually requires datasets that verify monotonic relationships among the samples. When the monotonic relationships are not met, changing the labels may be a viable option, but the risk is high: wrong label changes would completely change the information contained in the data. In this work, we tackle the construction of monotone datasets by removing the wrong or noisy examples that violate monotonicity restrictions. We propose two monotonic noise filtering algorithms to preprocess the ordinal datasets and improve the monotonic relations between instances. The experiments are carried out over eleven ordinal datasets, showing that the application of the proposed filters improve the prediction capabilities over different levels of noise.