Noise Elimination with a Re-Sampling Algorithm
Hugo Jair Escalante, Olac Fuentes · 2004
In this paper a new approach to noise detection and elimination in datasets for machine learning is presented. An algorithm that improves quality in training sets is introduced. This algorithm is based in the re-sampling idea that allows improving training data quality by identifying possible noisy instances and performing new measurements of each selected instance. We have obtained good results using this resampling algorithm in the prediction of stellar atmospheric parameters, a challenging astronomical domain where we tested the algorithm. We present experimental results of tests performed with varying noise levels that show how the re-sampling algorithm improves data quality, and hence classifier accuracy.