Stefan Todorov Hadjitodorov

Using diversity in cluster ensembles

The pairwise approach to cluster ensembles uses multiple partitions, each of which constructs a coincidence matrix between all pairs of objects. The matrices for the partitions are then combined and a final clustering is derived thereof. Here we study the diversity within such cluster ensembles. Based on this, we propose a variant of the generic ensemble method where the number of overproduced clusters is chosen randomly for every ensemble member (partition). Using three artificial sets we show that this approach increases the spread of the diversity within the ensemble thereby leading to a better match with the known cluster labels. Experimental results with three real data sets are also reported.

Moderate diversity for better cluster ensembles

Adjusted Rand index is used to measure diversity in cluster ensembles and a diversity measure is subsequently proposed. Although the measure was found to be related to the quality of the ensemble, this relationship appeared to be non-monotonic. In some cases, ensembles which exhibited a moderate level of diversity gave a more accurate clustering. Based on this, a procedure for building a cluster ensemble of a chosen type is proposed (assuming that an ensemble relies on one or more random parameters): generate a small random population of cluster ensembles, calculate the diversity of each ensemble and select the ensemble corresponding to the median diversity. We demonstrate the advantages of both our measure and procedure on 5 data sets and carry out statistical comparisons involving two diversity measures for cluster ensembles from the recent literature. An experiment with 9 data sets was also carried out to examine how the diversity-based selection procedure fares on ensembles of various sizes. For these experiments the classification accuracy was used as the performance criterion. The results suggest that selection by median diversity is no worse and in some cases is better than building and holding on to one ensemble.

Laryngeal pathology detection by means of class-specific neural maps

Most of the existing systems and methods for laryngeal pathology detection are characterized by a classification error. One of the basic problems is the approximation and estimation of the probability density functions of the given classes. In order to increase the accuracy of laryngeal pathology detection and to eliminate the most dangerous error classification of a patient with laryngeal disease as a normal speaker, an approach based on modeling of the probability density functions (pdfs) of the input vectors of the normal and pathological speakers by means of two prototype distribution maps (PDM), respectively, is proposed. The pdf of the input vectors of an unknown normal or pathological speaker is also modeled by such a prototype distribution neural map (PDM(X)), and the pathology detection is done by means of a ratio of specific similarities rather than by a direct comparison of some type of distance/similarity with a threshold. The experiments show an increased classification accuracy and that the proposed method can be used for screening the laryngeal diseases. The method is applied in a consulting system for clinical practice.

Fundamental frequency estimation of voice of patients with laryngeal disorders

The present paper is aimed at the development of new methods of fundamental frequency determination of voiced signal, uttered by patients with severe laryngeal disorders due to various diseases. The necessity to seek new methods is set by the increasing use of acoustic voice analysis as non-invasive technique for supporting diagnosis in laryngeal pathology. The classic methods, which do not take into account the peculiarities of the pathological voice, do not lead to satisfactory results in case of severely distorted periodicity of the signal. Three new methods are proposed in this paper for fundamental frequency determination: autocorrelation method, spectral method, and cepstral method. Comparison with the most commonly used methods is made.

Selecting diversifying heuristics for cluster ensembles

Cluster ensembles are deemed to be better than single clustering algorithms for discovering complex or noisy structures in data. Various heuristics for constructing such ensembles have been examined in the literature, e.g., random feature selection, weak clusterers, random projections, etc. Typically, one heuristic is picked at a time to construct the ensemble. To increase diversity of the ensemble, several heuristics may be applied together. However, not any combination may be beneficial. Here we apply a standard genetic algorithm (GA) to select from 7 standard heuristics for k-means cluster ensembles. The ensemble size is also encoded in the chromosome. In this way the data is forced to guide the selection of heuristics as well as the ensemble size. Eighteen moderate-size datasets were used: 4 artificial and 14 real. The results resonate with our previous findings in that high diversity is not necessarily a prerequisite for high accuracy of the ensemble. No particular combination of heuristics appeared to be consistently chosen across all datasets, which justifies the existing variety of cluster ensembles. Among the most often selected heuristics were random feature extraction, random feature selection and random number of clusters assigned for each ensemble member. Based on the experiments, we recommend that the current practice of using one or two heuristics for building k-means cluster ensembles should be revised in favour of using 3-5 heuristics.

Using key features in pattern classification

Abstract The term ‘key features’ is introduced to denote features with rare occurrence and great diagnostic power, which are implemented in a two-level classification scheme. Experiments with medical data illustrate the idea, confirming the expected improvement of the classification accuracy.

A two-level classifier for text-independent speaker identification

Abstract A two-level scheme for speaker identification is proposed. The first classifier level is based on the self-organizing map (SOM) of Kohonen. LPCC coefficients are used as input vectors for this classifier. LPCC coefficients are passed again through the already trained SOMs and as result the prototype distribution maps (PDMs) are obtained. The PDMs are the input for the second classifier level. The second level consists of multilayer perceptron (MLP) networks for each speaker. The first level of the classifier is a preprocessing procedure for the second level, where the final classification is made. The goal of the proposed approach is to combine the advantages of the two type of networks into one classification scheme in order to achieve higher identification accuracy. The experiments show an increased accuracy of the proposed two-level classifier, especially in the case of noise-corrupted signals.

A method for turbulent noise estimation in voiced signals.

AbstractIn this article a new acoustic parameter is introduced and it is shown that it may serve as an indicator of laryngeal function. It is termed the turbulent noise index (TNI) and is defined as 100(1−

An RBF network with tunable function shape

\bar R_{max}