Catálogo de publicaciones - libros

The analysis of gene expression data involves the observation of a very large number of variables (genes) on a few units (tissues). In such a context the recourse to conventional classification methods may be hard both for analytical and interpretative reasons. In this work a gene selection procedure for classification problems is addressed. The dimensionality reduction is based on the projections of genes along suitable directions obtained by Independent Factor Analysis (IFA). The performances of the proposed procedure are evaluated in the context of both supervised and unsupervised classification problems for different real data sets.

In this paper, a Monte Carlo study on the performance of two-mode cluster methods is presented. The synthetical data sets were generated to correspond to two types of data consisting of overlapping as well as disjoint clusters. Furthermore, the data sets differed in cluster number, degrees of within-group homogeneity and between-group heterogeneity as well as degree of cluster overlap. We found that the methods performed very differently depending on type of data, number of clusters, homogeneity and cluster overlap.

The problem of analysing interrater — agreement and — reliability is known both in human decision making and in machine interaction. Several measures have been developped in the last 100 years for this purpose, with Cohen’s Kappacoefficient to be the most popular one. Due to methodological considerations, the validity of kappa-type measures for interrater agreement has been discussed in a variety of papers. However, a global comparison of properties of these measures is currently still deficient. In our approach, we constructed an integral measure to evaluate the differences between two reliability measures for dichotomous ratings. Additionally, we studied bifurcation properties of the difference of these measures to quantify areas of minimal differences. From the methodological point of view, our integral-measure can also be used to construct other measures for interrater agreement.

In conventional multiclass classification learning, we seek to induce a prediction function from the domain of input patterns to a mutually exclusive set of class labels. As a straightforward generalization of this category of learning problems, so-called multi-label classification allows for input patterns to be associated with multiple class labels simultaneously. Text categorization is a domain of particular relevance which can be viewed as an instance of this setting. While the process of labeling input patterns for generating training sets already constitutes a major issue in conventional classification learning, it becomes an even more substantial matter of relevance in the more complex multi-label classification setting. We propose a novel active learning strategy for reducing the labeling effort and conduct an experimental study on the well-known Reuters-21578 text categorization benchmark dataset to demonstrate the efficiency of our approach.

In applications related to , the goal is not only to build a classifier for deciding whether a document among a list is relevant or not, but to learn a scoring function : → ℝ for ranking all possible documents with respect to their relevancy. Here we show how the boils down to binary classification with dependent data when accuracy is measured by the . The natural estimate of the risk being of the form of a , consistency of methods based on empirical risk minimization is studied using the theory of -processes. Taking advantage of this specific form, we prove that fast rates of convergence may be achieved under general noise assumptions.

In very many applications and, in particular, in PLS path modeling, it is of paramount importance to assess whether a set of variables is unidimensional. For this purpose, different methods are discussed. In addition to methods generally used in PLS path modeling, methods for the determination of the number of components in principal components analysis are considered. Two original methods based on permutation procedures are also proposed. The methods are compared to each others by means of a simulation study.

The PLS approach is a widely used technique to estimate path models relating various blocks of variables measured from the same population. It is frequently applied in the social sciences and in economics. In this type of applications, deviations from normality and outliers may occur, leading to an efficiency loss or even biased results. In the current paper, a robust path model estimation technique is being proposed, the partial robust M (PRM) approach. In an example its benefits are illustrated.

In this paper, a methodology is proposed which can be used for the identification of classes of units showing homogeneous behavioural models estimated through PLS Path Modelling. The proposed methodology aims at discovering or validating the existence of classes of units in PLS Path models in a predictive-oriented logic, such as it has been proposed, in the framework of PLS Regression, with PLS Typological Regression. An application to a study on customer satisfaction and loyalty is shown.

In many applications we are interested in finding clusters of data that share the same properties, like linear shape. We propose a hierarchical clustering procedure that merges groups if they are fitted well by the same linear model. The representative orthogonal model of each cluster is estimated robustly using iterated LQS regressions. We apply the method to two artificial datasets, providing a comparison of results against other non-hierarchical methods that can estimate linear clusters.

Multivariate outlier identification is often based on robust location and scatter estimates and usually performed relative to an elliptically shaped distribution. On the other hand, the idea of outlying observations is closely related to the notion of data depth, where observations with minimum depth are potential outliers. Here, we are not generally bound to the idea of an elliptical shape of the underlying distribution. Koshevoy and Mosler (1997) introduced zonoid trimmed regions which define a data depth. Recently, Paris Scholz (2002) and Becker and Paris Scholz (2004) investigated a new approach for robust estimation of convex bodies resulting from zonoids. We follow their approach and explore how the minimum volume zonoid (MZE) estimators can be used for multivariate outlier identification in the case of non-elliptically shaped null distributions.