Catálogo de publicaciones - libros

In this paper, we propose a novel filter for feature selection. Such filter relies on the estimation of the mutual information between features and classes. We bypass the estimation of the probability density function with the aid of the entropic-graphs approximation of Rényi entropy, and the subsequent approximation of the Shannon one. The complexity of such bypassing process does not depend on the number of dimensions but on the number of patterns/samples, and thus the curse of dimensionality is circumvented. We show that it is then possible to outperform a greedy algorithm based on the maximal relevance and minimal redundancy criterion. We successfully test our method both in the contexts of image classification and microarray data classification.

A successful interpretation of data goes through discovering crucial relationships between variables. Such a task can be accomplished by a Bayesian network. The dark side is that, when lots of variables are involved, the learning of the network slows down and may lead to wrong results. In this study, we demonstrate the feasibility of applying an existing Particle Swarm Optimization (PSO)-based approach to feature selection for filtering the irrelevant attributes of the dataset, resulting in a fine Bayesian network built with the K2 algorithm. Empirical tests carried out with real data coming from the bioinformatics domain bear out that the PSO fitness function is in a straight concordance to the most widely known validation measures for classification.

We present two feature selection methods, inspired in the Shannon’s entropy and the Information Gain measures, that are easy to implement. These methods apply when we have a database with continuous attributes and discrete multi- class. The first method applies when attributes are independent among them given the class. The second method is useful when we suspect that interdependencies among the attributes exist. In the experiments that we realized, with synthetic and real databases, the proposed methods are shown to be fast and to produce near optimum solutions, with a good feature reduction ratio.

This paper introduces an incremental data clustering algorithm based on the gravitational law. Basically, data samples are considered as unit-mass particles exposed to gravitational forces. Data points are clustered according their proximity during the simulation of the dynamical system defined by their gravitational fields. When the simulation is stopped, a set of prototypes is generated (several prototypes per cluster found). Each prototype will have associated a mass that is proportional to the number of particles in the sub-cluster and will be used as additional particle when new data samples are given for clustering. Experiments are performed on synthetic data sets and the obtained results are presented.

This paper analyzes, through clustering and visualization, Magnetic Resonance Spectra corresponding to a complex human brain tumour dataset. Clustering is performed as a two-stage process, in which the first stage model is Generative Topographic Mapping (GTM). In semi-supervised settings, class information can be added to refine the clustering process. A class information-enriched variant of GTM, class-GTM, is used here for a first cluster description of the data. The number of clusters used by GTM is usually large for visualization purposes and does not necessarily correspond to the overall class structure. Consequently, in a second stage, clusters are agglomerated using the K-means algorithm with different initialization strategies, some of them defined ad hoc for the GTM models. We aim to evaluate how and under what circumstances the use of class information influences tumour cluster-wise class separability in the final result of the two-stage clustering process.

Revealing the common underlying structure of data spread across multiple data sites by applying clustering techniques is the aim of collaborative clustering, a recent and innovative idea brought up on the basis of exchanging information granules instead of data patterns. The strength of the collaboration between each pair of data repositories is determined by a user-driven parameter, both in vertical and horizontal collaborative fuzzy clustering. In this study, Particle Swarm Optimization and Rough Set Theory are used for setting the most suitable values of the collaboration links between the data sites. Encouraging empirical results uncovered the deep impact observed at the individual clusters, allowing us to conclude that the overall effect of the collaboration has been improved.

One of the main factors for the knowledge discovery success is related to the comprehensibility of the patterns discovered by applying data mining techniques. Amongst which we can point out the Bayesian networks as one of the most prominent when considering the easiness of knowledge interpretation achieved. Bayesian networks, however, present limitations and disadvantages regarding their use and applicability. This paper presents an extension for the improvement of Bayesian networks, incorporating models of multiple regression for structure learning.

In electronic retailing market and/or supply chain, upstream sellers need to carefully plan and offer suitable returns policies to their downstream customers in order for ensuring product qualities, promoting product sales, maintaining customer satisfaction, lowering handling costs, and ultimately maximizing total profits. This paper aims at exploring the feasibility of using a hybrid data mining approach to support the coordination of returns policies and marketing plans for profit optimization in the e-business domain. A multi-dimensional data model and an integrated data mining process are provided to facilitate the three-staged clustering, classification, and association mining functions. Through this knowledge discovery process, customer and product classes identified in terms of the level of returns ratios are associated with the returns policies and marketing plans to generate rules for optimizing market profits. Also presented is a simulation example with embedded scenarios to test and validate the proposed data model and data mining process.

The wavelet transform has been used for feature extraction in many applications of pattern recognition. However, in general the learning algorithms are not designed taking into account the properties of the features obtained with discrete wavelet transform. In this work we propose a Markovian model to classify sequences of frames in the wavelet domain. The architecture is a composite of an external hidden Markov model in which the observation probabilities are provided by a set of hidden Markov trees. Training algorithms are developed for the composite model using the expectation-maximization framework. We also evaluate a novel delay-invariant representation to improve wavelet feature extraction for classification tasks. The proposed methods can be easily extended to model sequences of images. Here we present phoneme recognition experiments with TIMIT speech corpus. The robustness of the proposed architecture and learning method was tested by reducing the amount of training data to a few patterns. Recognition rates were better than those of hidden Markov models with observation densities based in Gaussian mixtures.

People that interact, cooperate or share common activities within information systems can be treated as a social network. The analysis of individual social standings appears to be a crucial element for the assessment of personal importance of each member within such weighted social network. The new measure of person significance – social position that depends on both the strength of relationships an individual maintains and social positions of all their acquaintances, together with its basic features and comparative experiments are presented in this paper.