Locally interpretable tree boosting: An application to house price prediction

Abstract

We introduce Locally Interpretable Tree Boosting (LitBoost), a tree boosting model tailored to applications where the data comes from several heterogeneous yet known groups with a limited number of observations per group. LitBoost constraints the complexity of a Gradient Boosted Trees model in a way that allows us to express the final model as a set of local Generalized Additive Models, yielding significant interpretability benefits while still maintaining some of the predictive power of a Gradient Boosted Trees model. We use house price prediction as a motivating example and demonstrate the performance of LitBoost on a data set of N=14382 observations from 15 different city districts in Oslo (Norway). We also test the robustness of LitBoost in an extensive simulation study on a synthetic data set.