GrMC: Towards Interpretable Classification Models That Are Also Accurate

Interpretability of classification models is an important issue. However, there is a lack of intrinsically interpretable models that are also highly accurate. To fill the gap, we introduce a new classification model type, namely Group Model Committee (GrMC), and an associated learning algorithm. Our key ideas are: (1) Divide a classification task's data space into several groups such that each group is defined by a simple condition and it has a unique, interpretable group model; (2) A data instance belongs to a group if it satisfies the group's defining condition and it is classified by utilizing the group models of the groups it belongs to. Experiments show that small interpretable GrMC models are often more accurate than existing intrinsically interpretable models, and also more accurate than Random Forests models. GrMC also has other strengths.