Cleaning Training-Datasets with Noise-Aware Algorithms

Abstract

We introduce a novel learning algorithm for noise elimination. Our algorithm is based on the re-measurement idea for the correction of erroneous observations and is able to discriminate between noisy and noiseless observations by using kernel methods. We apply our noise-aware algorithms to several domains including: astronomy, face recognition and ten machine learning benchmark datasets. Experimental results adding noise and useful anomalies to the data show that our algorithm improves data quality, without having to eliminate any observation from the original dataset