Enhancing Decision Tree Learning with Deep Networks

Prithaj Banerjee, Mahesh Lorik Yadav, Harish Guruprasad Ramaswamy, CHANDRA SHEKAR LAKSHMINARAYANAN

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Deep Learning, feature learning, oblique decision trees

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We devise a novel deep learning arcitecture that discovers features suitable for constructing oblique decision trees even in scenarios where greedy tree constructions fail.

Abstract: Conventional approaches to (oblique) decision tree construction for classification are greedy in nature. They can fail spectacularly when the true labeling function corresponds to a decision tree whose root node is uncorrelated with the labels (e.g. if the label function is the product of the sign of a collection of linear functions of the input). We define a new figure of merit to capture the usefulness of a linear function/hyperplane in a decision tree that is applicable even in scenarios where greedy procedures fail. We devise a novel deep neural network architecture that is very effective at seeking out hyperplanes/half-spaces/features that score highly on this metric. We exploit this property in a subroutine for a new decision tree construction algorithm. The proposed algorithm outperforms all other decision tree construction procedures, especially in situations where the hyper-planes corresponding to the top levels of the true decision tree are not useful features by themselves for classification but are essential for getting to full accuracy. The properties of the deep architecture that we exploit to construct the decision tree are also of independent interest, as they reveal the inner workings of the feature learning mechanism at play in deep neural networks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7201