Abstract:AdaBoost is one of the most popular ML algorithms. It is simple to implement and often found very effective by practitioners, while still being mathematically elegant and theoretically sound. AdaBoost's interesting behavior in practice still puzzles the ML community. We address the algorithm's stability and establish multiple convergence properties of "Optimal AdaBoost," a term coined by Rudin, Daubechies, and Schapire in 2004. We prove, in a reasonably strong computational sense, the almost universal existence of time averages, and with that, the convergence of the classifier itself, its generalization error, and its resulting margins, among many other objects, for fixed data sets under arguably reasonable conditions. Specifically, we frame Optimal AdaBoost as a dynamical system and, employing tools from ergodic theory, prove that, under a condition that Optimal AdaBoost does not have ties for best weak classifier eventually, a condition for which we provide empirical evidence from high dimensional real-world datasets, the algorithm's update behaves like a continuous map. We provide constructive proofs of several arbitrarily accurate approximations of Optimal AdaBoost; prove that they exhibit certain cycling behavior in finite time, and that the resulting dynamical system is ergodic; and establish sufficient conditions for the same to hold for the actual Optimal-AdaBoost update. We believe that our results provide reasonably strong evidence for the affirmative answer to two open conjectures, at least from a broad computational-theory perspective: AdaBoost always cycles and is an ergodic dynamical system. We present empirical evidence that cycles are hard to detect while time averages stabilize quickly. Our results ground future convergence-rate analysis and may help optimize generalization ability and alleviate a practitioner's burden of deciding how long to run the algorithm.

An empirical study of the convergence of regionboost

The Effect of Distance Metrics on Boosting with Dynamic Weighting Schemes

On the Convergence Properties of Optimal AdaBoost

AdaBoost in Region-Based Image Retrieval

A Convergence Rate Analysis for LogitBoost, MART and Their Variant.

High-Dimensional $L_2$Boosting: Rate of Convergence

The success of AdaBoost and its application in portfolio management

A new boosting algorithm based on dual averaging scheme

Boosting regression methods based on a geometric conversion approach: Using SVMs base learners

On the Margin Explanation of Boosting Algorithms.

TRBoost: a generic gradient boosting machine based on trust-region method

Dynamic Boosting in Deep Learning Using Reconstruction Error.

Feature Learning Viewpoint of AdaBoost and a New Algorithm

A Refined Margin Analysis for Boosting Algorithms Via Equilibrium Margin

Further Results on the Margin Explanation of Boosting: New Algorithm and Experiments

The Rate of Convergence of AdaBoost

Logistic Regression, AdaBoost and Bregman Distances

Boosting and Margin Theory

Boosting Through Optimization of Margin Distributions

W-Boost and Its Application to Web Image Classification

A Study of AdaBoost with Naive Bayesian Classifiers: Weakness and Improvement.