Abstract:Margin theory provides one of the most popular explanations to the success of AdaBoost, where the central point lies in the recognition that margin is the key for characterizing the performance of AdaBoost. This theory has been very influential, e.g., it has been used to argue that AdaBoost usually does not overfit since it tends to enlarge the margin even after the training error reaches zero. Previously the minimum margin bound was established for AdaBoost, however, Breiman (1999) [9] pointed out that maximizing the minimum margin does not necessarily lead to a better generalization. Later, Reyzin and Schapire (2006) [37] emphasized that the margin distribution rather than minimum margin is crucial to the performance of AdaBoost. In this paper, we first present the kth margin bound and further study on its relationship to previous work such as the minimum margin bound and Emargin bound. Then, we improve the previous empirical Bernstein bounds (Audibert et al. 2009; Maurer and Pontil, 2009) [2,30], and based on such findings, we defend the margin-based explanation against Breimanʼs doubts by proving a new generalization error bound that considers exactly the same factors as Schapire et al. (1998) [39] but is sharper than Breimanʼs (1999) [9] minimum margin bound. By incorporating factors such as average margin and variance, we present a generalization error bound that is heavily related to the whole margin distribution. We also provide margin distribution bounds for generalization error of voting classifiers in finite VC-dimension space.

Optimal Minimal Margin Maximization with Boosting

Boosting the Minimum Margin: LPBoost vs. AdaBoost

Boosting Through Optimization of Margin Distributions

A Refined Margin Analysis for Boosting Algorithms Via Equilibrium Margin

On the doubt about margin explanation of boosting

Margins are Insufficient for Explaining Gradient Boosting

Boosting and Margin Theory

Max-margin Based Bayesian Classifier.

On the Margin Explanation of Boosting Algorithms.

The Many Faces of Optimal Weak-to-Strong Learning

LPBoost with Strong Classifiers

Selective Boosting Algorithm for Maximizing the Soft Margin

The Kth, Median and Average Margin Bounds for AdaBoost

Analysis of boosting algorithms using the smooth margin function

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

A Duality View of Boosting Algorithms

On the Dual Formulation of Boosting Algorithms