Algorithm Selection for Classification Problems via Cluster-based Meta-features

Meta-features describe the characteristics of the datasets to facilitate algorithm selection. This paper proposes a new set of meta-features based on clustering the instances within datasets. We propose the use of a greedy clustering algorithm, and evaluate the meta-features generated based on the learning curves produced by the Random Forest algorithm. We also compared the utility of the proposed meta-features against preexisting meta-features described in the literature, and evaluated the applicability of these meta-features over a sample of UCI datasets. Our results show that these meta-features do indeed improve the performance when applied to the algorithm selection task.