Catálogo de publicaciones - libros

Determining the relevant features is a combinatorial task in various fields of machine learning such as text mining, bioinformatics, pattern recognition, etc. Several scholars have developed various methods to extract the relevant features but no method is really superior. Breiman proposed Random Forest to classify a pattern based on CART tree algorithm and his method turns out good results compared to other classifiers. Taking advantages of Random Forest and using wrapper approach which was first introduced by Kohavi , we propose an algorithm named Dynamic Recursive Feature Elimination (DRFE) to find the optimal subset of features for reducing noise of the data and increasing the performance of classifiers. In our method, we use Random Forest as induced classifier and develop our own defined feature elimination function by adding extra terms to the feature scoring. We conducted experiments with two public datasets: Colon cancer and Leukemia cancer. The experimental results of the real world data showed that the proposed method has higher prediction rate compared to the baseline algorithm. The obtained results are comparable and sometimes have better performance than the widely used classification methods in the same literature of feature selection.

There have been a great deal of research on learning from imbalanced datasets. Among the widely used methods proposed to solve such a problem, the most common are based either on under or over sampling of the original dataset. In this work, we evaluate several methods of under-sampling, such as Tomek Links, with the goal of improving the performance of the classifiers generated by different ML algorithms (decision trees, support vector machines, among others) applied to problem of determining the structural similarity of proteins.

Prediction of protein interactions is one of the central problems in post–genomic biology. In this paper, we present an association rule-based protein interaction prediction method. We adopted neural network to cluster protein interaction data, and used information theory based feature selection method to reduce protein feature dimension. After model training, feature association rules are generated to interaction prediction by decoding a set of learned weights of trained neural network and by mining association rules. For model training, an initial network model was constructed with public protein interaction data considering their functional categories, set of features, and interaction partners. The prediction performance was compared with traditional simple association rule mining method. The experimental results show that proposed method has about 96.1% interaction prediction accuracy compared to simple association mining approach which achieved about 91.4% accuracy.

This paper presents the use of neural networks for the prediction of protein Secondary Structure. We propose a pre-processing stage based on the method of Cascaded Nonlinear Components Analysis (C-NLPCA), in order to get a dimensional reduction of the data which may consider its nonlinearity. Then, the reduced data are placed in predictor networks and its results are combined. For the verification of possible improvements brought by the use of C-NLPCA, a set of tests was done and the results will be demonstrated in this paper. The C-NLPCA revealed to be efficient, propelling a new field of research.

Because the lengths of nucleotide sequences for microorganisms are various, it is difficult to directly compare the complete nucleotide sequences among microorganisms. In this study, we adopted a method that can convert DNA sequences of microorganisms into numerical form then applied Fourier transform to the numerical DNA sequences in order to investigate the distributions of nucleotides. Also, a visualization scheme for transformed DNA sequences was proposed to help visually categorize microorganisms. Furthermore, the well-known neural network technique Self-Organizing Map (SOM) was applied to the transformed DNA sequences to draw conclusions of taxonomic relationships among the bacteria of Bacillus genus. The results show that the relationships among the bacteria are corresponding to recent biological findings.

In this paper, the kurtosis-based method for the classification of mental activities is proposed. The EEG signals were recorded during imagination of left or right hand movement. The kurtosis of EEG and its dynamic properties with respect to time are analyzed. The experiment results show that the kurtosis can reflect the EEG pattern changes of different motor imageries. According to the analysis and experiment results, a kurtosis based classifier for the classification of left and right movement imagination is designed. This classifier can achieves near 90% correct rate. As the kurtosis is computationally less demanding and can also be estimated in on-line way, so the new method proposed in this paper has the practicability in the application of brain-computer interface.

Previous works showed that the joint use of Principal Component Analysis (PCA) and Independent Component Analysis (ICA) allows to extract a few meaningful dominant components from the EEG of patients in coma. A procedure for automatic critical epoch detection might support the doctor in the long time monitoring of the patients, this is why we are headed to find a procedure able to automatically quantify how much an epoch is critical or not. In this paper we propose a procedure based on the extraction of some features from the dominant components: the entropy and the kurtosis. This feature analysis allowed us to detect some epochs that are likely to be critical and that are worth inspecting by the expert in order to assess the possible restarting of the brain activity.

Diffusion Tensor (DT) fiber tracking techniques offer significant potential for studying anatomical connectivity of human brain . And the reliability and accuracy of fiber tracking results depend on the quality of estimated DT which is determined by parameters of image acquisition protocol. The aim of this paper is to investigate what echo-planar image (EPI) acquisition parameters: the number of sensitizing directions K and diffusion weighting b-value gives the best estimation of DT and shorter scan time. We carried out tracking on synthetic dataset that was artificially corrupted by various levels of Gaussian noise to study the effects of K and b-value on fiber tracking results, and to evaluate the quality of estimated DT. It was found that when K value larger than 13 and b-value larger than 800 smm best estimated DTs. And further increments of K and b-value had no significant effect on quality of estimated DT.

The papers shows, through theoretical studies and simulations, that using the description of the plant by Takagi-Sugeno (T-S), it is possible to design a nonlinear controller to control the position of the leg of a paraplegic patient. The control system was designed to change the angle of the joint knee of 60°. This is the first study that describes the application of Takagi-Sugeno (T-S) models in this kind of problem.

Presence of mass in breast tissues is highly indicative of breast cancer. The research work investigates the significance of neural-association of mass type of breast abnormality patterns for benign and malignant class characterization using auto-associator neural network and original features. The characterized patterns are finally classified into benign and malignant classes using a classifier neural network. Grey-level based statistical features, BI-RADS features, patient age feature and subtlety value feature have been used in proposed research work. The proposed research technique attained a 94% testing classification rate with a 100% training classification rate on digital mammograms taken from the DDSM benchmark database.