Methods for voting classification algorithms, such as Bagging and AdaBoost, have been shown to be very successful in improving the accuracy of certain classifiers for artificial and real-world datasets. We review these algorithms and describe a large empirical study comparing several variants in conjunction with a decision tree inducer (three variants) and a Naive-Bayes inducer. The purpose of the study is to improve our understanding of why and when these algorithms, which use perturbation, reweighting, and combination techniques, affect classification error. We provide a bias and variance decomposition of the error to show how different methods and variants influence these two terms. This allowed us to determine that Bagging reduced variance of unstable methods, while boosting methods (AdaBoost and Arc-x4) reduced both the bias and variance of unstable methods but increased the variance for Naive-Bayes, which was very stable. We observed that Arc-x4 behaves differently than AdaBoost if reweighting is used instead of resampling, indicating a fundamental difference. Voting variants, some of which are introduced in this paper, include: pruning versus no pruning, use of probabilistic estimates, weight perturbations (Wagging), and backfitting of data. We found that Bagging improves when probabilistic estimates in conjunction with no-pruning are used, as well as when the data was backfit. We measure tree sizes and show an interesting positive correlation between the increase in the average tree size in AdaBoost trials and its success in reducing the error. We compare the mean-squared error of voting methods to non-voting methods and show that the voting methods lead to large and significant reductions in the mean-squared errors. Practical problems that arise in implementing boosting algorithms are explored, including numerical instabilities and underflows. We use scatterplots that graphically show how AdaBoost reweights instances, emphasizing not only "hard" areas but also outliers and noise.

Research in machine learning, statistics and related fields has produced a wide variety of algorithms for classification. However, most of these algorithms assume that all errors have the same cost, which is seldom the case in KDD prob- lems. Individually making each classification learner costsensitive is laborious, and often non-trivial. In this paper we propose a principled method for making an arbitrary classifier cost-sensitive by wrapping a cost-minimizing procedure around it. This procedure, called MetaCost, treats the underlying classifier as a black box, requiring no knowledge of its functioning or change to it. Unlike stratification, MetaCost is applicable to any number of classes and to arbitrary cost matrices. Empirical trials on a large suite of benchmark databases show that MetaCost almost always produces large cost reductions compared to the cost-blind classifier used (C4.5RULES) and to two forms of stratification. Further tests identify the key components of MetaCost and those that can be varied without substantial loss. Experiments on a larger database indicate that MetaCost scales well.

Tree induction is one of the most effective and widely used methods for building classification models. However, many applications require cases to be ranked by the probability of class membership. Probability estimation trees (PETs) have the same attractive features as classification trees (e.g., comprehensibility, accuracy and efficiency in high dimensions and on large data sets). Unfortunately, decision trees have been found to provide poor probability estimates. Several techniques have been proposed to build more accurate PETs, but, to our knowledge, there has not been a systematic experimental analysis of which techniques actually improve the probability-based rankings, and by how much. In this paper we first discuss why the decision-tree representation is not intrinsically inadequate for probability estimation. Inaccurate probabilities are partially the result of decision-tree induction algorithms that focus on maximizing classification accuracy and minimizing tree size (for example via reduced-error pruning). Larger trees can be better for probability estimation, even if the extra size is superfluous for accuracy maximization. We then present the results of a comprehensive set of experiments, testing some straghtforward methods for improving probability-based rankings. We show that using a simple, common smoothing method--the Laplace correction--uniformly improves probability-based rankings. In addition, bagging substantioJly improves the rankings, and is even more effective for this purpose than for improving accuracy. We conclude that PETs, with these simple modifications, should be considered when rankings based on class-membership probability are required.

Abstract. Many KDD systems incorporate an implicit or explicit preference for simpler models, but this use of “Occam’s razor ” has been strongly criticized by several authors (e.g., Schaffer, 1993; Webb, 1996). This controversy arises partly because Occam’s razor has been interpreted in two quite different ways. The first interpretation (simplicity is a goal in itself) is essentially correct, but is at heart a preference for more comprehensible models. The second interpretation (simplicity leads to greater accuracy) is much more problematic. A critical review of the theoretical arguments for and against it shows that it is unfounded as a universal principle, and demonstrably false. A review of empirical evidence shows that it also fails as a practical heuristic. This article argues that its continued use in KDD risks causing significant opportunities to be missed, and should therefore be restricted to the comparatively few applications where it is appropriate. The article proposes and reviews the use of domain constraints as an alternative for avoiding overfitting, and examines possible methods for handling the accuracy–comprehensibility trade-off.

Decision trees are one of the most effective and widely used classification methods. However, many applications require class probability estimates, and probability estimation trees (PETs) have the same attractive features as classification trees (e.g., comprehensibility, accuracy and efficiency in high dimensions and on large data sets). Unfortunately, decision trees have been found to provide poor probability estimates. Several techniques have been proposed to build more accurate PETs, but, to our knowledge, there has not been a systematic experimental analysis of which techniques actually improve the probability estimates, and by how much. In this paper we first discuss why the decision-tree representation is not intrinsically inadequate for probability estimation. Inaccurate probabilities are partially the result of decision-tree induction algorithms that focus on maximizing classification accuracy and minimizing tree size (for example via reduced-error pruning). Larger tree...

We study the issue of error diversity in ensembles of neural networks. In ensembles of regression estimators, the measurement of diversity can be formalised as the Bias-Variance-Covariance decomposition. In ensembles of classifiers, there is no neat theory in the literature to date. Our objective is to understand how to precisely define, measure, and create diverse errors for both cases. As a focal point we study one algorithm, Negative Correlation (NC) Learning which claimed, and showed empirical evidence, to enforce useful error diversity, creating neural network ensembles with very competitive performance on both classification and regression problems. With the lack of a solid understanding of its dynamics, we engage in a theoretical and empirical investigation. In an initial empirical stage, we demonstrate the application of an evolutionary search algorithm to locate the optimal value for λ, the configurable parameter in NC. We observe the behaviour of the optimal parameter under different ensemble architectures and datasets; we note a high degree of unpredictability, and embark on a more formal investigation. During the theoretical investigations, we find that NC succeeds due to exploiting the

Occam's razor has been the subject of much controversy. This paper argues that this is partly because it has been interpreted in two quite different ways, the first of which (simplicity is a goal in itself) is essentially correct, while the second (simplicity leads to greater accuracy) is not. The paper reviews the large variety of theoretical arguments and empirical evidence for and against the "second razor," and concludes that the balance is strongly against it. In particular, it builds on the case of (Schaffer, 1993) and (Webb, 1996) by considering additional theoretical arguments and recent empirical evidence that the second razor fails in most domains. A version of the first razor more appropriate to KDD is proposed, and we argue that continuing to apply the second razor risks causing significant opportunities to be missed. 1 Occam's Two Razors William of Occam's famous razor states that "Nunquam ponenda est pluralitas sin necesitate," which, approximately translated, means "En...

Current methods to avoid overfitting are either data-oriented (using separate data for validation) or representation-oriented (penalizing complexity in the model). This paper proposes process-oriented evaluation, where a model's expected generalization error is computed as a function of the search process that led to it. The paper develops the necessary theoretical framework, and applies it to one type of learning: rule induction. A process-oriented version of the CN2 rule learner is empirically compared with the default CN2. The process-oriented version is more accurate in a large majority of the datasets, with high significance, and also produces simpler models. Experiments in artificial domains suggest that processoriented evaluation is particularly useful in high-dimensional domains. 1 INTRODUCTION Overfitting avoidance is often considered the central problem of machine learning (e.g., (Cheeseman & Oldford, 1994)). If a learner is sufficiently powerful, it must guard against selec...

Classifier committee learning methods generate multiple classifiers to form a committee by repeated application of a single base learning algorithm. The committee members vote to decide the final classification. Two such methods, Bagging and Boosting, have shown great success with decision tree learning. They create different classifiers by modifying the distribution of the training set. This paper studies a different approach: Stochastic Attribute Selection Committee learning of decision trees. It generates classifier committees by stochastically modifying the set of attributes but keeping the distribution of the training set unchanged. An empirical evaluation of a variant of this method, namely Sasc, in a representative collection of natural domains shows that the SASC method can significantly reduce the error rate of decision tree learning. On average Sasc is more accurate than Bagging and less accurate than Boosting, although a one-tailed sign-test fails to show that these differences are significant at a level of 0.05. In addition, it is found that, like Bagging, Sasc is more stable than Boosting in terms of less frequently obtaining significantly higher error rates than C4.5 and, when error is raised, producing lower error rate increases. Moreover, like Bagging, Sasc is amenable to parallel and distributed processing while Boosting is not.

Bagging constructs an estimator by averaging predictors trained on bootstrap samples. Bagged estimates almost consistently improve on the original predictor. It is thus important to understand the reasons for this success, and also for the occasional failures. It is widely believed that bagging is effective thanks to the variance reduction stemming from averaging predictors. However, seven years from its introduction, bagging is still not fully understood. This paper provides experimental evidence supporting the hypothesis that bagging stabilizes prediction by equalizing the influence of training examples. This eect is detailed in two dierent frameworks: estimation on the real line and regression. Bagging's improvements/deteriorations are explained by the goodness/badness of highly influential examples, in situations where the usual variance reduction argument is at best questionable. Finally, reasons for the equalization effect are advanced. They support that other resampling strategies such as half-sampling should provide qualitatively identical effects while being computationally less demanding than bootstrap sampling.