On Computing Optimal Tree Ensembles

TL;DR

This paper introduces novel algorithms for computing optimal decision tree ensembles, improving tractability results and providing bounds, with potential practical implications for classification and regression tasks.

Contribution

It presents two new algorithms for optimal tree ensembles, including a witness-tree technique and dynamic programming approach, advancing the computational methods for these models.

Findings

Algorithms for optimal tree ensembles with improved runtime bounds.

Dynamic programming approach for computing ensembles.

Ensembles may require exponentially fewer cuts than single trees.

Abstract

Random forests and, more generally, (decision\nobreakdash-)tree ensembles are widely used methods for classification and regression. Recent algorithmic advances allow to compute decision trees that are optimal for various measures such as their size or depth. We are not aware of such research for tree ensembles and aim to contribute to this area. Mainly, we provide two novel algorithms and corresponding lower bounds. First, we are able to carry over and substantially improve on tractability results for decision trees: We obtain an algorithm that, given a training-data set and an size bound $S \in \mathbb{R}$, computes a tree ensemble of size at most $S$ that classifies the data correctly. The algorithm runs in $(4\delta D S)^S \cdot poly$-time, where $D$ the largest domain size, $\delta$ is the largest number of features in which two examples differ, $n$ the number of input examples, and $poly$ a polynomial of the input size. For decision trees, that is, ensembles of size 1, we obtain a running time of $(\delta D s)^s \cdot poly$, where $s$ is the size of the tree. To obtain these algorithms, we introduce the witness-tree technique, which seems promising for practical implementations. Secondly, we show that dynamic programming, which has been applied successfully to computing decision trees, may also be viable for tree ensembles, providing an $\ell^n \cdot poly$-time algorithm, where $\ell$ is the number of trees. Finally, we compare the number of cuts necessary to classify training data sets for decision trees and tree ensembles, showing that ensembles may need exponentially fewer cuts for increasing number of trees.