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.

Analysis of Classification Margin for Classification Accuracy with Applications

On the Optimization of Margin Distribution

Max-margin Based Bayesian Classifier.

Margin Distribution Analysis

A Multi-Class Large Margin Classifier

Theory and Analysis of Double-Margin SVM

On the doubt about margin explanation of boosting

Negative Margin Matters: Understanding Margin in Few-Shot Classification

Boosting Through Optimization of Margin Distributions

A Refined Margin Analysis for Boosting Algorithms Via Equilibrium Margin

Adjustable entropy funtion method for misclassification minimization problems

Nonlinear Online Classification Algorithm with Probability Margin

Improved Sample Complexities for Deep Networks and Robust Classification via an All-Layer Margin

Unified Binary and Multiclass Margin-Based Classification

Mistake, Manipulation and Margin Guarantees in Online Strategic Classification

Optimal Minimal Margin Maximization with Boosting

Comparative Analysis on Margin Based Feature Selection Algorithms.

Discriminative Learning Method Based on Minimum Classification Error Criterion

Margin attribute reductions for multi-label classification

Large-margin Feature Selection for Monotonic Classification

Coherence Functions for Multicategory Margin-based Classification Methods