Neural Regression Trees

TL;DR

This paper introduces a neural regression tree model that jointly learns optimal discretization thresholds and feature representations, improving regression performance on challenging tasks.

Contribution

It presents a novel neural regression tree framework that optimizes discretization and features simultaneously, surpassing existing RvC methods.

Findings

Achieved state-of-the-art results on two regression benchmarks.

Demonstrated the effectiveness of joint optimization in RvC.

Validated the model's superiority over ad-hoc discretization approaches.

Abstract

Regression-via-Classification (RvC) is the process of converting a regression problem to a classification one. Current approaches for RvC use ad-hoc discretization strategies and are suboptimal. We propose a neural regression tree model for RvC. In this model, we employ a joint optimization framework where we learn optimal discretization thresholds while simultaneously optimizing the features for each node in the tree. We empirically show the validity of our model by testing it on two challenging regression tasks where we establish the state of the art.