Learning with Local Models

Next to prediction accuracy, the interpretability of models is one of the fundamental criteria for machine learning algorithms. While high accuracy learners have intensively been explored, interpretability still poses a difficult problem, largely because it can hardly be formalized in a general way. To circumvent this problem, one can often find a model in a hypothesis space that the user regards as understandable or minimize a user-defined measure of complexity, such that the obtained model describes the essential part of the data. To find interesting parts of the data, unsupervised learning has defined the task of detecting local patterns and subgroup discovery. In this paper, the problem of detecting local classification models is formalized. A multi-classifier algorithm is presented that finds a global model that essentially describes the data, can be used with almost any kind of base learner and still provides an interpretable combined model.