. This paper shows that the accuracy of learned text classifiers can be improved by augmenting a small number of labeled training documents with a large pool of unlabeled documents. This is important because in many text classification problems obtaining training labels is expensive, while large quantities of unlabeled documents are readily available. We introduce an algorithm for learning from labeled and unlabeled documents based on the combination of Expectation-Maximization (EM) and a naive Bayes classifier. The algorithm first trains a classifier using the available labeled documents, and probabilistically labels the unlabeled documents. It then trains a new classifier using the labels for all the documents, and iterates to convergence. This basic EM procedure works well when the data conform to the generative assumptions of the model. However these assumptions are often violated in practice, and poor performance can result. We present two extensions to the algorithm that improve ...

One key difficulty with text classification learning algorithms is that they require many hand-labeled examples to learn accurately. This dissertation demonstrates that supervised learning algorithms that use a small number of labeled examples and many inexpensive unlabeled examples can create high-accuracy text classifiers. By assuming that documents are created by a parametric generative model, Expectation-Maximization (EM) finds local maximum a posteriori models and classifiers from all the data -- labeled and unlabeled. These generative models do not capture all the intricacies of text; however on some domains this technique substantially improves classification accuracy, especially when labeled data are sparse. Two problems arise from this basic approach. First, unlabeled data can hurt performance in domains where the generative modeling assumptions are too strongly violated. In this case the assumptions can be made more representative in two ways: by modeling sub-topic class structure, and by modeling super-topic hierarchical class relationships. By doing so, model probability and classification accuracy come into correspondence, allowing unlabeled data to improve classification performance. The second problem is that even with a representative model, the improvements given by unlabeled data do not sufficiently compensate for a paucity of labeled data. Here, limited labeled data provide EM initializations that lead to low-probability models. Performance can be significantly improved by using active learning to select high-quality initializations, and by using alternatives to EM that avoid low-probability local maxima.

In this paper we perform a comparison among FSS-EBNA, a randomized, populationbased and evolutionary algorithm, and two genetic and other two sequential search approaches in the well known Feature Subset Selection (FSS) problem. In FSS-EBNA, the FSS problem, stated as a search problem, uses the EBNA (Estimation of Bayesian Network Algorithm) search engine, an algorithm within the EDA (Estimation of Distribution Algorithm) approach. The EDA paradigm is born from the roots of the GA community in order to explicitly discover the relationships among the features of the problem and not disrupt them by genetic recombination operators. The EDA paradigm avoids the use of recombination operators and it guarantees the evolution of the population of solutions and the discovery of these relationships by the factorization of the probability distribution of best individuals in each generation of the search. In EBNA, this factorization is carried out by a Bayesian network induced by a chea...

Receiver Operating Characteristic (ROC) curves are a standard way to display the performance of a set of binary classifiers for all feasible ratios of the costs associated with false positives and false negatives. For linear classifiers, the set of classifiers is typically obtained by training once, holding constant the estimated slope and then varying the intercept to obtain a parameterized set of classifiers whose performances can be plotted in the ROC plane. We consider the alternative of varying the asymmetry of the cost function used for training. We show that the ROC curve obtained by varying both the intercept and the asymmetry, and hence the slope, always outperforms the ROC curve obtained by varying only the intercept. In addition, we present a path-following algorithm for the support vector machine (SVM) that can compute efficiently the entire ROC curve, and that has the same computational complexity as training a single classifier. Finally, we provide a theoretical analysis of the relationship between the asymmetric cost model assumed when training a classifier and the cost model assumed in applying the classifier. In particular, we show that the mismatch between the step function used for testing and its convex upper bounds, usually used for training, leads to a provable and quantifiable difference around extreme asymmetries.

Methods to avoid overfitting fall into two broad categories: data-oriented (using separate data for validation) and representation-oriented (penalizing complexity in the model). Both have limitations that are hard to overcome. We argue that fully adequate model evaluation is only possible if the search process by which models are obtained is also taken into account. To this end, we recently proposed a method for process-oriented evaluation (POE), and successfully applied it to rule induction [ Domingos, 1998b ] . However, for the sake of simplicity this treatment made a number of rather artificial assumptions. In this paper the assumptions are removed, and a simple formula for error estimation is obtained. Empirical trials show the new, better-founded form of POE to be as accurate as the previous one, while further reducing theory sizes. 1 Introduction Overfitting avoidance is a central problem in machine learning. If a learner is su#ciently powerful, whatever repre...

Recent studies (Alizadeh et al. 2000, Bittner et al. 2000, Golub et al. 1999) demonstrate the discovery of disease subtypes from gene expression data. In this paper, we propose a principled and systematic approach to address the computational problem of partitioning the set of sample tissues into statistically meaningful classes. We start by describing a method, called overabundance analysis, for assessing how informative a given expression data set is with respect to a partition of the samples. As we show, in several published expression datasets, an overabundance of genes separating known classes is observed. Then, we use this method as the foundation to a novel approach to class discovery. In this approach, we search for partitions that have statistically significant overabundance score. We evaluate the performance of our approach on synthetic data, where we show it can recover planted partitions. Finally, we apply it to several published tumor expression datasets, and show that we find several highly pronounced partitions.

There has been an explosive growth of data-mining models involving latent structure for clustering and classification. While having related objectives these models use different parameterizations and often very different specifications and constraints. Model choice is thus a major methodological issue and a crucial practical one for applications. In this paper, we work from a general formulation of hierarchical Bayesian mixed-membership models in Erosheva [15] and Erosheva, Fienberg, and Lafferty [19] and present several model specifications and variations, both parametric and nonparametric, in the context of the learning the number of latent groups and associated patterns for clustering units. Model choice is an issue within specifications, and becomes a component of the larger issue of model comparison. We elucidate strategies for comparing models and specifications by producing novel analyses of two data sets: (1) a corpus of scientific publications from the Proceedings of the National Academy of Sciences (PNAS) examined earlier by Erosheva, Fienberg, and Lafferty [19] and Griffiths and Steyvers [22]; (2) data on functionally disabled American seniors from the National

Using multiple regularization hyperparameters is an effective method for managing model complexity in problems where input features have varying amounts of noise. While algorithms for choosing multiple hyperparameters are often used in neural networks and support vector machines, they are not common in structured prediction tasks, such as sequence labeling or parsing. In this paper, we consider the problem of learning regularization hyperparameters for log-linear models, a class of probabilistic models for structured prediction tasks which includes conditional random fields (CRFs). Using an implicit differentiation trick, we derive an efficient gradient-based method for learning Gaussian regularization priors with multiple hyperparameters. In both simulations and the real-world task of computational RNA secondary structure prediction, we find that multiple hyperparameter learning provides a significant boost in accuracy compared to models learned using only a single regularization hyperparameter. 1

In order to select a good hypothesis language (or model) from a collection of possible models, one has to assess the generalization performance of the hypothesis which is returned by a learner that is bound to use some particular model. This paper deals with a new and very efficient way of assessing this generalization performance. We present a new analysis which characterizes the expected generalization error of the hypothesis with least training error in terms of the distribution of error rates of the hypotheses in the model. This distribution can be estimated very efficiently from the data which immediately leads to an efficient model selection algorithm. The analysis predicts learning curves with a very high precision and thus contributes to a better understanding of why and when over-fitting occurs. We present empirical studies (controlled experiments on Boolean decision trees and a large-scale text categorization problem) which show that the model selection algorithm leads to err...

Model selection [e.g., 1] is considered the problem of choosing a hypothesis language which provides an optimal balance between low empirical error and high structural complexity. In this Abstract, we discuss the intuition of a new, very efficient approach to model selection. Our approach is inherently Bayesian [e.g., 2], but instead of using priors on target functions or hypotheses, we talk about priors on error values -- which leads us to a new mathematical characterization of the expected true error. In the setting of classification learning, a learner is given a sample, drawn according to an unknown distribution of labeled instances, and returns the empirical minimizer (the hypothesis with the least empirical error) which has a certain (unknown) true error. If this process is carried out repeatedly, the true error of the empirical minimizer will vary from run to run as the empirical minimizer depends on the (randomly drawn) sample. This induces a distribution of true errors of empirical minimizers, over the possible samples drawn according to the unknown distribution. If this distribution would be known, one could easily derive the expected true error of the empirical minimizer of a model by integrating over this distribution. This would immediately lead to an optimal model selection algorithm: Enumerate the models, calculate the expected error of each model by integrating over the error distribution, and select the model with the least expected error. PAC theory [3] and the VC framework provide worst-case bounds on the chance of drawing a sample such that the true error of the minimizer exceeds some " -- "worst-case" meaning that they hold for any distribution of instances and any concept in a given class. By contrast, we focus on how to determine this distributi...