Meta-Learning Approach for Noise Filter Algorithm Recommendation

Abstract

Preprocessing techniques can increase the quality or even enable Machine Learning algorithms. However, it is not simple to identify the preprocessing algorithms we should apply. This work proposes a methodology to recommend a noise filtering algorithm based on Meta-Learning, predicting which algorithm should be chosen based on a set of features calculated from a dataset. From synthetics datasets, we created the meta-data from an extracted set of meta-features and the f1-score performance metric calculated from the DT, KNN, and RF classifiers. To perform the suggestion, we used a meta-ranker that returns the rank of the best algorithms. We selected three noise filtering algorithms, HARF, GE, and ORBoost. To predict the f1-score, we used the PCT, RF, and KNN algorithms as meta-rankers. Our results indicate that the proposed solution acquired over 60% and 80% accuracy when considering a top-1 and top-2 approach. It also shows that the meta-rankers, when compared with a random choice and single algorithms as a baseline, provided an overall performance gain for the Machine Learning algorithm.