Active Dataset Generation for Meta-learning System Quality Improvement

Meta-learning use meta-features to formally describe datasets and find possible dependencies of algorithm performance from them. But there is not enough of various datasets to fill a meta-feature space with acceptable density for future algorithm performance prediction. To solve this problem we can use active learning. But it is required ability to generate nontrivial datasets that can help to improve the quality of the meta-learning system. In this paper we experimentally compare several such approaches based on maximize diversity and Bayesian optimization.