Affiliation:
1. University of Zagreb Faculty of Organization and Informatics, Pavlinska 2, Varaždin
Abstract
Automated Machine Learning (AutoML) utilizing meta-learning (M-L) has gained prominence in the scientific community. Current M-L methods necessitate substantial data and computational resources for extracting meta-features encoding data properties. However, the time needed for meta-feature extraction exceeds that for predictions in M-L systems. This article proposes a domain-specific M-L paradigm tailored to social science, aiming to identify universally applicable meta-features in social science data. Investigating domain-specific properties, the study discerned common meta-features across social science domains, facilitating an efficient AutoML strategy with reduced data requirements. Ninety meta-features, clustered into eight groups characterizing social science data, were employed, focusing on education and business domains. An analysis of 46 datasets revealed domain-specific variations in meta-feature values, confirmed by Wilcoxon tests. Notably, certain meta-features exhibited consistency across social science domains, demonstrating potential for cross-domain AutoML adoption. This research introduces a targeted M-L approach, optimizing AutoML efficiency for social science applications by identifying common meta-features across diverse domains.
Funder
Hrvatska Zaklada za Znanost
Publisher
Association for Information Communication Technology Education and Science (UIKTEN)