Olga Georgieva

Dynamic data assigning assessment clustering of streaming data

Discovering interesting patterns or substructures in data streams is an important challenge in data mining. Clustering algorithms are very often applied to identify single substructures although they are designed to partition a data set. Another problem of clustering algorithms is that most of them are not designed for data streams. This paper discusses a recently introduced procedure that deals with both problems. The procedure explores ideas from cluster analysis, but was designed to identify single clusters without the necessity to partition the whole data set into clusters. The new extended version of the algorithm is an incremental clustering approach applicable to stream data. It identifies new clusters formed by the incoming data and updates the data space partition. Clustering of artificial and real data sets illustrates the abilities of the proposed method.

Takagi-Sugeno fuzzy model development of batch biotechnological processes

Abstract The paper deals with the modelling of batch biotechnological processes based on the Takagi–Sugeno (TS) fuzzy model. Two possible process descriptions namely the input–output TS fuzzy model and state-space TS fuzzy model are discussed. A parameterised input–output TS fuzzy model is proposed. The model identification is based on the extended structure and parameter identification procedure by means of product-space fuzzy clustering. It compensates the insensibility of the clustering procedure with respect to the parameters that are maintained at a constant level during the batch fermentation. Parameterised input–output TS fuzzy model and state-space TS fuzzy model of xanthan gum production by strain Xanthomonas campestris IST-342 are introduced as a meter of illustration.

Gustafson-Kessel algorithm for evolving data stream clustering

A simplified clustering algorithm that enables on-line partitioning of data streams is proposed. The algorithm applies adaptive-distance metric to identify clusters with different shape and orientation. It is applicable to a wide range of practical evolving system type applications as diagnostics and prognostics, system identification, real time classification, and process quality monitoring and control.

Learning to decode human emotions from event-related potentials

Abstract Reported works on electroencephalogram (EEG)-based emotion recognition systems generally employ the principles of supervised learning to build subject-dependent (single/intra-subject) models. Building subject-independent (multiple/inter-subject) models is a harder problem due to the EEG data variability between subjects. The contribution of this paper is twofold. First, we provide a framework for selection of a small number of basic temporal features, event-related potential (ERP) amplitudes, and latencies that are sufficiently robust to discriminate emotion states across multiple subjects. Second, we test comparatively the feasibility of six standard unsupervised (clustering) techniques to build intra-subject and inter-subject models to discriminate emotion valence in the ERPs collected while subjects were viewing high arousal images with positive or negative emotional content.

Software reliability assessment via fuzzy logic model

Recently in software engineering it became of great importance to be able to reason about non functional characteristics of software. This holds for a large variety of application areas like embedded and safety-critical software systems as well as service-oriented systems. Currently there exist a number of models for evaluation of software reliability, which are based on statistics and probability theory. However, the assumptions of existing models introduce imprecision in the reliability estimations. In this paper we are making a step towards solving of this problem by introducing a fuzzy logic approach for estimation of software reliability. Our approach is further validated with a case-study.

Stability of quasilinear fuzzy system

Abstract This paper deals with the stability of a system described by a quasilinear fuzzy model. The stability analysis is considered by investigation of the transition process of the system. Two cases are discussed — when all subsystems are stable and when one or more subsystems are unstable. A non-conventional decision of stabilization of a locally unstable system is proposed.

Cluster analysis for EEG biosignal discrimination

The paper aims to define the ability of unsupervised learning approach to identify emotional biosignals evoked while viewing affected pictures. Two problems are consequently resolved. First, the most important features of the Electroencephalography (EEG) data set have been selected. Secondly, cluster analysis technique is applied in order to extract the specific knowledge of the existing dependencies. The clustering results of particular data subsets are presented and discussed.

Cluster validity measures based on the minimum description length principle

Determining the number of clusters is a crucial problem in cluster analysis. Cluster validity measures are one way to try to find the optimum number of clusters, especially for prototype-based clustering. However, no validity measure turns out to work well in all cases. In this paper, we propose an approach to determine the number of cluster based on the minimum description length principle which does not need high computational costs and is also applicable in the context of fuzzy clustering.

Fuzzy Clustering of Macroarray Data

The complete sequence of bacterial genomes provides new perspectives for the study of gene expression and gene function. DNA array experiments allow measuring the expression levels for all genes of an organism in a single hybridization experiment.

Fuzzy models of rainfall: discharge dynamics

Three different methods for building Takagi-Sugeno models relating rainfall to catchment discharge are tested on the Zwalm catchment. They correspond to the following identification methods: Grid Partitioning (GP), Subtractive Clustering (SC), and Gustafson-Kessel clustering (GK). The models are parametrized on a one-year identification data set and tested against the complete five-year data set. Although these models show a similar behaviour, resulting in comparable values of the Nash and Suttcliffe criterion and the root mean square error, the best values are obtained for the models generated using the GK method.