Complexity Measurement for Multitask Classification Problems in Machine Learning

Multitask classification problems are common in many areas of machine learning. Intrinsic measures of the complexity of the target labels and classes within the heterogenous problem space are crucial as they allow for a preliminary evaluation of the problem's difficulty, as well as an estimation of how suitable different machine learning methods might be for tackling these problems. Prior works in developing complexity measures for classification problems have mostly focused on single and multi-label classification problems. In this paper, we propose five novel data quality measures that assess the complexity of multitask problems. These measures, rooted in Rough Set Theory and principles of granular computing, can manage hybrid data (encompassing both numeric and categorical elements) and incomplete datasets. We demonstrate the effectiveness and practical utility of these proposed measures by conducting experiments on several public datasets, thereby providing tangible evidence of their potential in handling real-world scenarios.