In the feature subset selection problem, a learning algorithm is faced with the problem of selecting a relevant subset of features upon which to focus its attention, while ignoring the rest. To achieve the best possible performance with a particular learning algorithm on a particular training set, a feature subset selection method should consider how the algorithm and the training set interact. We explore the relation between optimal feature subset selection and relevance. Our wrapper method searches for an optimal feature subset tailored to a particular algorithm and a domain. We study the strengths and weaknesses of the wrapper approach and show a series of improved designs. We compare the wrapper approach to induction without feature subset selection and to Relief, a filter approach to feature subset selection. Significant improvement in accuracy is achieved for some datasets for the two families of induction algorithms used: decision trees and

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.

In this doctoral dissertation, we study three basic problems in machine learning and two new hypothesis spaces with corresponding learning algorithms. The problems we investigate are: accuracy estimation, feature subset selection, and parameter tuning. The latter two problems are related and are studied under the wrapper approach. The hypothesis spaces we investigate are: decision tables with a default majority rule (DTMs) and oblivious read-once decision graphs (OODGs).

The simple Bayesian classifier (SBC), sometimes called Naive-Bayes, is built based on a conditional independence model of each attribute given the class. The model was previously shown to be surprisingly robust to obvious violations of this independence assumption, yielding accurate classification models even when there are clear conditional dependencies. The SBC can serve as an excellent tool for initial exploratory data analysis when coupled with a visualizer that makes its structure comprehensible. We describe such a visual representation of the SBC model that has been successfully implemented. We describe the requirements we had for such a visualization and the design decisions we made to satisfy them. Keywords:Classification, simple/naive-Bayes, visualization.

We consider feature selection in the "wrapper " model of feature selection. This typically involves an NP-hard optimization problem that is approximated by heuristic search for a "good" feature subset. First considering the idealization where this optimization is performed exactly, we give a rigorous bound for generalization error under feature selection. The search heuristics typically used are then immediately seen as trying to achieve the error given in our bounds, and succeeding to the extent that they succeed in solving the optimization. The bound suggests that, in the presence of many "irrelevant" features, the main source of error in wrapper model feature selection is from "overfitting " hold-out or cross-validation data. This motivates a new algorithm that, again under the idealization of performing search exactly, has sample complexity (and error) that grows logarithmically in the number of "irrelevant" features -- which means it can tolerate having a number of "irrelevant" f...

Most classification algorithms receive as input a set of attributes of the classified objects. In many cases, however, the supplied set of attributes is not sufficient for creating an accurate, succinct and comprehensible representation of the target concept. To overcome this problem, researchers have proposed algorithms for automatic construction of features. The majority of these algorithms use a limited predefined set of operators for building new features. In this paper we propose a generalized and flexible framework that is capable of generating features from any given set of constructor functions. These can be domain-independent functions such as arithmetic and logic operators, or domain-dependent operators that rely on partial knowledge on the part of the user. The paper describes an algorithm which receives as input a set of classified objects, a set of attributes, and a specification for a set of constructor functions that contains their domains, ranges and properties. The algorithm produces as output a set of generated features that can be used by standard concept learners to create improved classifiers. The algorithm maintains a set of its best generated features and improves this set iteratively. During each iteration, the algorithm performs a beam search over its defined feature space and constructs new features by applying constructor functions to the members of its current feature set. The search is guided by general heuristic measures that are not confined to a specific feature representation. The algorithm was applied to a variety of classification problems and was able to generate features that were strongly related to the underlying target concepts. These features also significantly improved the accuracy achieved by standard concept learners, for a ...

: Each year, one of the explicit challenges for the KDD research community is to develop methods that facilitate the use of inductive learning algorithms for mining very large databases. By collecting, categorizing, and summarizing past work on scaling up inductive learning algorithms, this paper serves to establish a common ground for researchers addressing the challenge. We begin with a discussion of important, but often tacit, issues related to scaling up learning algorithms. We highlight similarities among methods by categorizing them into three main approaches. For each approach, we then describe, compare, and contrast the different constituent methods, drawing on specific examples from the published literature. Finally, we use the preceding analysis to suggest how one should proceed when dealing with a large problem, and where future research efforts should be focused. Primary contact: Foster Provost NYNEX Science and Technology, 400 Westchester Avenue, White Plains, NY 10604 em...

In this paper, we propose a dynamic approach to feature and classifier selection. In our approach, based on performance, visual features and classifiers are selected automatically. In earlier work, we presented the Visual Apprentice, in which users can define visual object models via a multiple-level object definition hierarchy (region, perceptual-area, object-part, and object). Visual Object Detectors are learned, using various learning algorithms- as the user provides examples from images or video, visual features are extracted and multiple classifiers are learned for each node of the hierarchy. In this paper, features and classifiers are selected automatically at each node, depending on their performance over the training set provided by the user. Thus, changes in the training data yield dynamic changes in the features and classifiers used. We introduce the concept of Recurrent Visual Semantics and show how it can be used to identify domains in which performancebased learning techni...

In this chapter I begin by discussing Distributed Data Mining (DDM) for scaling up, beginning by asking what scaling up means, questioning whether it is necessary, and then presenting a brief survey of what has been done to date. I then provide motivation beyond scaling up, arguing that DDM is a more natural way to view data mining generally. DDM eliminates many difficulties encountered when coalescing already-distributed data for monolithic data mining, such as those associated with heterogeneity of data and with privacy restrictions. By viewing data mining as inherently distributed, important open research issues come into focus, issues that currently are obscured by the lack of explicit treatment of the process of producing monolithic data sets. I close with a discussion of the necessity of DDM for an efficient process of knowledge discovery.

This paper establishes common ground for researchers addressing the challenge of scaling up inductive data mining algorithms to very large databases, and for practitioners who want to understand the state of the art. We begin with a discussion of important, but often tacit, issues related to scaling up. We then overview existing methods, categorizing them into three main approaches. Finally, we use the overview to recommend how to proceed when dealing with a large problem and where future research efforts should be focused.