Abstract
In Data Science, synthetic datasets allow comparing algorithms under controlled conditions. In multi-label classification, current dataset generators ignore label associations, which is relevant in areas such as medicine to represent comorbidities, for instance. This work proposes a generator that explicitly incorporates these associations. Each instance is created by two mechanisms: (i) label generation using conditional co-occurrence probabilities; (ii) feature generation with normal distributions whose mean depends on the present labels. Its differential is enabling algorithm analysis and comparison while simultaneously controlling label correlation and their separability in the feature space.

This publication has 5 references indexed in Scilit: