Coherent Collections of Rules Describing Exceptional Materials Identified with a Multi-Objective Optimization of Subgroups

Using a modest amount of data from a large population, subgroup discovery (SGD) identifies outstanding subsets of data with respect to a certain property of interest of that population. The SGs are described by "rules". These are constraints on key descriptive parameters that characterize the material or the environment. These parameters and constraints are obtained by maximizing a quality function that establishes a tradeoff between SG size and utility, i.e., between generality and exceptionality. The utility function measures how outstanding a SG is. However, this approach does not give a unique solution, but typically many SGs have similar quality-function values. Here, we identify coherent collections of SGs of a "Pareto region" presenting various size-utility tradeoffs and define a SG similarity measure based on the Jaccard index, which allows us to hierarchically cluster these optimal SGs. These concepts are demonstrated by learning rules that describe perovskites with high bulk modulus. We show that SGs focusing on exceptional materials exhibit a high quality-function value but do not necessarily maximize it. We compare the mean shift with the cumulative Jensen-Shannon divergence (D_sJS) as utility functions and show that the SG rules obtained with D_cJS are more focused than those obtained with the mean shift.