Catálogo de publicaciones - libros

Many processes experience abrupt changes in their dynamics. This causes problems for some prediction algorithms which assume that the dynamics of the sequence to be predicted are constant, or at least only change slowly over time. In this paper the problem of predicting sequences with sudden changes in dynamics is considered. For a model of multivariate Gaussian data we derive expected generalization error of standard linear Fisher classifier in situation where after unexpected task change, the classification algorithm learns on a mixture of old and new data. We show both analytically and by an experiment that optimal length of learning sequence depends on complexity of the task, input dimensionality, on the power and periodicity of the changes. The proposed solution is to consider a collection of agents, in this case non-linear single layer perceptrons (agents), trained by a memetic like learning algorithm. The most successful agents are voting for predictions. A grouped structure of the agent population assists in obtaining favorable diversity in the agent population. Efficiency of socially organized evolving multi-agent system is demonstrated on an artificial problem.

We present a method, called equivalence learning, which applies a two-class classification approach to object-pairs defined within a multi-class scenario. The underlying idea is that instead of classifying objects into their respective classes, we classify object pairs either as equivalent (belonging to the same class) or non-equivalent (belonging to different classes). The method is based on a vectorisation of the similarity between the objects and the application of a machine learning algorithm (SVM, ANN, LogReg, Random Forests) to learn the differences between equivalent and non-equivalent object pairs, and define a unique kernel function that can be obtained via equivalence learning. Using a small dataset of archaeal, bacterial and eukaryotic 3-phosphoglycerate-kinase sequences we found that the classification performance of equivalence learning slightly exceeds those of several simple machine learning algorithms at the price of a minimal increase in time and space requirements.

Algorithms for text classification generally involve two stages, the first of which aims to identify textual elements (words and/or phrases) that may be relevant to the classification process. This stage often involves an analysis of the text that is both language-specific and possibly domain-specific, and may also be computationally costly. In this paper we examine a number of alternative keyword-generation methods and phrase-construction strategies that identify key words and phrases by simple, language-independent statistical properties. We present results that demonstrate that these methods can produce good classification accuracy, with the best results being obtained using a phrase-based approach.

We address the problem of learning automatically to map heterogeneous semi-structured documents onto a mediated target XML schema. We adopt a machine learning approach where the mapping between input and target documents is learned from a training corpus of documents. We first introduce a general stochastic model of semi structured documents generation and transformation. This model relies on the concept of meta-document which is a latent variable providing a link between input and target documents. It allows us to learn the correspondences when the input documents are expressed in a large variety of schemas. We then detail an instance of the general model for the particular task of HTML to XML conversion. This instance is tested on three different corpora using two different inference methods: a dynamic programming method and an approximate LaSO-based method.

Fractal theory has been used for computer graphics, image compression and different fields of pattern recognition. In this paper, a fractal based method for recognition of both on-line and off-line Farsi/ Arabic handwritten digits is proposed. Our main goal is to verify whether fractal theory is able to capture discriminatory information from digits for pattern recognition task. Digit classification problem (on-line and off-line) deals with patterns which do not have complex structure. So, a general purpose fractal coder, introduced for image compression, is simplified to be utilized for this application. In order to do that, during the coding process, contrast and luminosity information of each point in the input pattern are ignored. Therefore, this approach can deal with on-line data and binary images of handwritten Farsi digits. In fact, our system represents the shape of the input pattern by searching for a set of geometrical relationship between parts of it. Some fractal-based features are directly extracted by the fractal coder. We show that the resulting features have invariant properties which can be used for object recognition.

Description logics have emerged as one of the most successful formalisms for knowledge representation and reasoning. They are now widely used as a basis for ontologies in the Semantic Web. To extend and analyse ontologies, automated methods for knowledge acquisition and mining are being sought for. Despite its importance for knowledge engineers, the learning problem in description logics has not been investigated as deeply as its counterpart for logic programs.

We propose the novel idea of applying evolutionary inspired methods to solve this task. In particular, we show how Genetic Programming can be applied to the learning problem in description logics and combine it with techniques from Inductive Logic Programming. We base our algorithm on thorough theoretical foundations and present a preliminary evaluation.

This paper addresses relation information extraction problem and proposes a method of discovering relations among entities which is buried in different nest structures of XML documents. The method first identifies and collects XML fragments that contain all types of entities given by users, then computes similarity between fragments based on semantics of their tags and their structures, and clusters fragments by similarity so that the fragments containing the same relation are clustered together, finally extracts relation instances and patterns of their occurrences from each cluster. The results of experiments show that the method can identify and extract relation information among given types of entities correctly from all kinds of XML documents with meaningful tags.