Catálogo de publicaciones - libros

Sammon’s mapping is a well-known procedure for mapping data from a higher-dimensional space onto a lower-dimensional one. But the original algorithm has a disadvantage. It lacks generalization, which means that new points cannot be added to the obtained map without recalculating it. The SAMANN neural network, that realizes Sammon’s algorithm, provides a generalization capability of projecting new data. A drawback of using SAMANN is that the training process is extremely slow. One of the ways of speeding up the neural network training process is to use parallel computing. In this paper, we proposed some parallel realizations of the SAMANN.

Center selection based on proper orthogonal decomposition (POD) is presented to select centers for the radial basis function (RBF) neural network in prediction of nonlinear time series. The proposed method takes advantages of the time-sequence feature in time series data and enables the center selection to be implemented in a parallel manner. Simulations on a benchmark problem and on two predictions of stock prices show that the presented method can be applied effectively to the prediction of nonlinear time series. Besides possessing higher precisions in training and testing, the proposed method has stronger generalization and noise resistance abilities, compared to several other popular center selection methods.

This paper proposes the Spectral Clustering Kernel Machine (SCKM) for times series prediction. Support Vector Machine (SVM), Relevance Vector Machine (RVM) and the Spectral Clustering Kernel Machine (SCKM) are compared in terms of performance accuracy for a simple time series approximation problem. The three outlined algorithms each of which with interesting features to perform automated learning are examined, analysed and empirically tested. In case of the SVM, our tests combine also a preprocessing stage including Kohonen Maps (SOM) as well as K-means clustering. In the case of RVM we also implemented a constructive approach based on the fast marginal likelihood maximization described in [14]. Prediction results in two benchmark time series have been addressed using various performance metrics. The results demonstrate that whereas RVM models achieve larger parsimony of the fitted model, both SVM and SCKM attain higher accuracy. The learning models are competitive for real world problems.

In this paper, we explore the possibilities of support vector machines to identify peer-to-peer (p2p) traffic in high-performance routers with packet sampling. Commercial networks limit user access bandwidth -either physically or logically-. However, in research networks there are no bandwidth restrictions, since this would interfere with research tasks. User behavior in research networks has changed radically with the advent of p2p multimedia file transfers: many users take advantage of the huge bandwidth (e.g. compared to domestic DSL access) to exchange movies and the like. This behavior may have a deep impact on research network utilization. Consequently, in the framework of the MOLDEIP project, we have proposed to apply support vector machine detection to identify those activities in high-performance research network routers. Due to their high port rates, those routers cannot extract the headers of all the packets that traverse them, but only a sample. The results in this paper suggest that support vector machine detection of p2p traffic in high-performance routers with packet sampling is highly successful and outperforms recent approaches like [1].

Standard kernel-based classifiers use only a single kernel, but the real-world applications and the recent developments of various kernel methods have emphasized the need to consider a combination of multiple kernels. We propose an evolutionary approach for finding the optimal weights of a combined kernel used by the Support Vector Machines (SVM) algorithm for solving some particular problems. We use a genetic algorithm (GA) for evolving these weights. The numerical experiments show that the evolved combined kernels (ECKs) perform better than the convex combined kernels (CCKs) for several classification problems.

Relevance Vector Machines (RVM) extend Support Vector Machines (SVM) to have probabilistic interpretations, to build sparse training models with fewer basis functions (i.e., relevance vectors or prototypes), and to realize Bayesian learning by placing priors over parameters (i.e., introducing hyperparameters). However, RVM algorithms do not scale up to large data sets. To overcome this problem, in this paper we propose a RVM boosting algorithm and demonstrate its potential with a text mining application. The idea is to build weaker classifiers, and then improve overall accuracy by using a boosting technique for document classification. The algorithm proposed is able to incorporate all the training data available; when combined with sampling techniques for choosing the working set, the boosted learning machine is able to attain high accuracy. Experiments on REUTERS benchmark show that the results achieve competitive accuracy against state-of-the-art SVM; meanwhile, the sparser solution found allows real-time implementations.

The use of support vector machines for multi-category problems is still an open field to research. Most of the published works use the one-against-rest strategy, but with a one-against-one approach results can be improved. To avoid testing with all the binary classifiers there are some methods like the Decision Directed Acyclic Graph based on a decision tree. In this work we propose an optimization method to improve the performance of the binary classifiers using Particle Swarm Optimization and an automatic method to build the graph that improves the average number of operations needed in the test phase. Results show a good behavior when both ideas are used.

In the study, a new dynamic fuzzy model is proposed in combination with support vector machine (SVM) to explore stock market dynamism. The fuzzy model integrates various factors with influential degree as the input variables, and the genetic algorithm (GA) adjusts the influential degree of each input variable dynamically. SVM then serves to predict stock market dynamism in the next phase. In the meanwhile, the multiperiod experiment method is designed to simulate the volatility of stock market. Then, we compare it with other methods. The model from the study does generate better results than others.

The quality of rubber compounds is assessed by rheological and mechanical tests. Since mechanical tests are very time consuming, the main idea of this work is to quest for strong nonlinear relationships between rheological and mechanical tests in order to reduce the latter. The multilayered perceptron and support vector machine combined with data preprocessing were applied to model hardness and density of the vulcanizates from the rheological parameters of the raw compounds. The results outline the advantage of proper data preprocessing.

In this study, a multistage evolutionary programming (EP) based support vector machine (SVM) ensemble model is proposed for designing a corporate bankruptcy prediction system to discriminate healthful firms from bad ones. In the proposed model, a bagging sampling technique is first used to generate different training sets. Based on the different training sets, some different SVM models with different parameters are then trained to formulate different classifiers. Finally, these different SVM classifiers are aggregated into an ensemble output using an EP approach. For illustration, the proposed SVM ensemble model is applied to a real-world corporate failure prediction problem.