Catálogo de publicaciones - libros

This paper introduces a novel architecture to efficiently code in a self-organized manner, data from sequences or a hierarchy of sequences. The main objective of the architecture proposed is to achieve an inductive model of the sequential data through a learning algorithm in a finite vector space with generalization and prediction properties improved by the compression process. The architecture consists of a hierarchy of recurrent self-organized maps with emergence which performs a fractal codification of the sequences. An adaptive outlier detection algorithm is used to automatically extract the emergent properties of the maps. A visualization technique to help the analysis and interpretation of data is also developed. Experiments and results for the architecture are shown for an anomaly intrusion detection problem.

Density based clustering methods allow the identification of arbitrary, not necessarily convex regions of data points that are densely populated. The number of clusters does not need to be specified beforehand; a cluster is defined to be a connected region that exceeds a given density threshold. This paper introduces the notion of local scaling in density based clustering, which determines the density threshold based on the local statistics of the data. The local maxima of density are discovered using a -nearest-neighbor density estimation and used as centers of potential clusters. Each cluster is grown until the density falls below a pre-specified ratio of the center point’s density. The resulting clustering technique is able to identify clusters of arbitrary shape on noisy backgrounds that contain significant density gradients. The focus of this paper is to automate the process of clustering by making use of the local density information for arbitrarily sized, shaped, located, and numbered clusters. The performance of the new algorithm is promising as it is demonstrated on a number of synthetic datasets and images for a wide range of its parameters.

A system of rule extraction out of a complex hierarchical classifier is proposed in this paper. There are several methods for rule extraction out of trained artificial neural networks (ANN’s), but these methods do not scale well, results are satisfactory for small problems. For complicated problems hundreds of rules are produced, which are hard to govern.

In this paper a hierarchical classifier with a tree-like structure and simple ANN’s at nodes, is presented, which splits the original problem into several sub-problems that overlap. Node classifiers are all ( with accuracy only better than random), and errors are corrected at lower levels. Single sub-problems constitute of examples that were hard to separate. Such architecture is able to classify better than single network models.

At the same time rules are extracted, which only answer which sub-problem a given example belongs to. Such rules, by introducing hierarchy, are simpler and easier to modify by hand, giving also a better insight into the original classifier behaviour.

The purpose of this work is to demonstrate that it is possible to cluster contact maps for pairs of alpha helices such that each of the clusters corresponds to a group of pairs of alpha helices with similar properties. The property of the configuration of helix pairs that was chosen for study is the packing attribute. The contact maps are compared to one another using a novel contact map comparison scheme based upon the locations of contacts in the contact maps. A k-nearest neighbours technique is used to perform the clustering, and the cosine between vectors corresponding to contact map regions was the distance metric. The clustering of contact maps to determine whether maps corresponding to similar packing values are placed into the same clusters yielded promising results.

In this paper we look at a dynamic method for analysing data, called a data probe. The probe flies through the data space and is affected by the proximity and number of data points. The trajectory followed by the probe provides information about how the data are organised geometrically. We apply a state feedback method to the probe equations to make the probe search out certain data values.

The problem of classifying chemical compounds is studied in this paper. An approach based on minimal contrast and common topological patterns discovered from compounds dataset is presented. The algorithm is strongly associated with the classical emerging patterns techniques known from decision tables. We tested the proposed algorithm on real classification problems. Results show that it provides better classification accuracy than other existing algorithms. Another advantage of the introduced classifier is that it has a simple, understandable structure and can be easily extended by the expert knowledge.

This paper deals with power-quality (PQ) event detection, classification and characterization using higher-order sliding cumulants to examine the signals. Their maxima and minima are the main features, and the classification strategy is based in competitive layers. Concretely, we concentrate on the task of differentiating two types of transients (short duration and long duration). By measuring the fourth-order central cumulants’ maxima and minima, we build the two-dimensional feature measured vector. Cumulants are calculated over high-pass digitally filtered signals, to avoid the low-frequency 50-Hz signal. We have observed that the minima and maxima measurements produce clusters in the feature space for th-order cumulants; rd-order cumulants are not capable of differentiate these two very similar PQ events. The experience aims to set the foundations of an automatic procedure for PQ event detection.

We focus on the recently introduced nearest neighbor based entropy estimator from Kraskov, Stögbauer and Grassberger (KSG) [10], the nearest neighbor search of which is performed by the so called box assisted algorithm [7]. We compare the performance of KSG with respect to three spatial indexing methods: box-assisted, -D trie and projection method, on a problem of mutual information estimation of a variety of pdfs and dimensionalities. We conclude that the -D trie method is significantly faster then box-assisted search in fixed-mass and fixed-radius neighborhood searches in higher dimensions. The projection method is much slower than both alternatives and not recommended for practical use.

Identifying bacterial promoters is an important step towards understanding gene regulation. In this paper, we address the problem of predicting the location of promoters in . Language of bacterial sequence can be described using formal system such a context-free grammar, and problem of promoter region recognition replaced by grammar induction. The accepted method for this problem is to use grammar-based classifier system (GCS).

In this paper we propose a new method for learning Bayesian network classifiers in an indirect way instead of directly from data. This new model is a classifier based on [1] that is a probabilistic graphical model similar to Bayesian networks but in which directed cycles are allowed. The benefits from doing things in this way are that learning process for dependency networks can be easier and simpler than learning Bayesian networks, with the direct consequence that learning algorithms could have good properties about scalability. We show that it is possible to take advantage of this facility to get Bayesian networks classifiers without losing quality in classification.