Sparse regression over clusters: SparClur

Abstract

Prediction tasks in personalized medicine require models that combine accuracy and interpretability. We propose an integer optimization approach for building sparse regression models with enforced coordination, using data partitioned among leaves in a prediction tree. We show that the method recovers the true underlying relationship between observations and target variables in large-scale synthetic data in seconds. We apply our method to several real-world medical prediction problems and observe that the additional structure imposed provides a substantial gain in interpretability, at a low cost to accuracy.