Decision trees for heterogeneous dose-response signal analysis

We propose a novel decision tree algorithm for modeling function-valued responses. This algorithm partitions the feature space into homogeneous subpopulations with common dose-response signals using a splitting criterion based on Nadaraya-Watson kernel regression and the Cramér-von Mises statistical test. We formulate an important business problem of sales team composition within the dose-response framework. Experimental results on generated and real-world sales data show the efficacy of the approach.