Towards automatic generation of metafeatures

Abstract

The selection of metafeatures for metalearning (MtL) is often an ad hoc process. The lack of a proper motivation for the choice of a metafeature rather than others is questionable and may originate a loss of valuable information for a given problem (e.g., use of class entropy and not attribute entropy). We present a framework to systematically generate metafeatures in the context of MtL. This framework decomposes a metafeature into three components: meta-function, object and postprocessing. The automatic generation of metafeatures is triggered by the selection of a meta-function used to systematically generate metafeatures from all possible combinations of object and post-processing alternatives. We executed experiments by addressing the problem of algorithm selection in classification datasets. Results show that the sets of systematic metafeatures generated from our framework are more informative than the non-systematic ones and the set regarded as state-of-the-art.