Feature construction using explanations of individual predictions

Feature construction can contribute to comprehensibility and performance of machine learning models. Unfortunately, it usually requires exhaustive search in the attribute space or time-consuming human involvement generate meaningful features. We propose a novel heuristic approach for reducing based on aggregation instance-based explanations predictive The proposed Explainable Construction (EFC) methodology identifies groups co-occurring attributes exposed by popular explanation methods, such as IME SHAP. empirically show that these significantly reduces time feature using logical, relational, Cartesian, numerical, threshold num-of-N X-of-N constructive operators. An analysis 10 transparent synthetic datasets shows EFC effectively informative constructs relevant Using 30 real-world classification datasets, we significant improvements accuracy several classifiers demonstrate feasibility even large datasets. Finally, generated interpretable features problem from financial industry, which were confirmed domain expert.