Datasets meta-feature description for recommending feature selection algorithm

Abstract

Meta-learning is an approach for solving the algorithm selection problem, which is how to choose the best algorithm for a certain task. This task corresponds to a dataset in machine learning and data mining. The main challenge in meta-learning is to engineer a meta-feature description for datasets. In the paper we apply meta-learning for feature selection. We found a meta-feature set which showed the best result in predicting proper feature selection algorithms. We also suggested a novel approach to engineer meta-features for data preprocessing algorithms, which is based on estimating the best parametrization of processing algorithms on small subsamples.