Search

This research deals with dimensionality reduction technique which is based on radial basis function (RBF) theory. The technique uses RBF for mapping multidimensional data points into a low-dimensional space by interpolating the previously calculated position of so-called control points. This paper analyses various ways of selection of control points (regularized orthogonal least squares method, random and stratified selections). The experiments have been carried out with 8 real and artificial data sets. Positions of the control points in a low-dimensional space are found by principal component analysis. We demonstrate that random and stratified selections of control points are efficient and acceptable in terms of balance between projection error (stress) and time-consumption.DOI: 10.15181/csat.v4i1.1095

This research deals with projection error evaluation for large data sets using only a personal computer without any particular technologies for high performance computing. A shortcoming of basic projection error calculation ways is such that they require a large amount of computer memory or computation time is not acceptable when large data sets are analyzed. This paper proposes two ways for projection error evaluation: the first one is based on calculating the projection error for not full data set, but only for representative data sample, the second one obtains the projection error by dividing a data set into the smaller data sets. The experiments have been carried out with twelve real and artificial data sets. The computational efficiency of the projection error evaluation ways is confirmed by a comprehensive set of comparisons. We demonstrate that dividing data set into the smaller data sets allows us to calculate the projection error for large data sets.

In this paper, the projection evaluation measures such as stress function, Spearman’s rho, Konig’s topology preservation, silhouette and Renyi entropy have been analyzed. The principal component analysis (PCA) and part–linear multidimensional projection (PLMP) techniques are used to reduce the dimensionality of the initial data set. The experiments have been carried out with seven real and artificial datasets. The experimental investigation has shown that several quality evaluation measures have to be used when dimension reduction problem is solved.

Tobulėjant šiuolaikinėms informacinėms ir komunikacinėms technologijoms, sparčiai didėja apdorojamų ir saugomų duomenų kiekiai, todėl duomenų analizės uždavinys tampa vis sudėtingesnis, sunku daryti greitus, efektyvius ir teisingus sprendimus. Duomenų analizei dažnai pasitelkiama duomenų tyryba. Duomenų tyryba – tai procesas, kurio metu iš duomenų išgaunamos naudingos žinios. Duomenims apdoroti bei žinioms išgauti reikalingos duomenų tyrybos sistemos, leidžiančios apdoroti įvairios apimties duomenis. Tyrime siekiama nustatyti, kokios apimties duomenis per priimtiną laiką sugeba apdoroti populiariausios duomenų tyrybos sistemos. Nagrinėjamas ir lyginamas trijose atvirojo kodo duomenų tyrybos sistemose (WEKA, KNIME, ORANGE) įgyvendintų klasifi kavimo ir klasterizavimo algoritmų skaičiavimo laikas, analizuojant skirtingos apimties duomenų aibes. Vertinant sistemas svarbus ne tik algoritmų skaičiavimo laikas, bet ir klasifi kavimo bei klasterizavimo tikslumas, kurį pavyksta pasiekti per tą laiką, todėl straipsnyje pateikiamos ir eksperimentiniuose tyrimuose gauto tikslumo matų reikšmės. Investigation of the abilities of data mining systems to analyse various volume datasets Kotryna Paulauskienė, Olga Kurasova Summary The aim of the paper is to determine what volume of data the popular data mining systems are able to analyse within a reasonable period of time, when solving classifi cation and clustering problems. Three open source data mining systems are investigated: WEKA, KNIME, and ORANGE. The experiments have been carried out with eight datasets, where the number of attributes was fi xed – 100 and the number of instances ranged between 5000 and 600 000. The experimental investigation has shown that when the ORANGE system is used, the data of more than 50 000 instances are of too large volume. In order to analyse larger datasets, the WEKA and KNIME systems need to be used. The data of more than 200 000 instances are of too large volume for WEKA and KNIME, however, when simple classifi cation methods are used, both systems are able to handle 400 000 instances, and KNIME – 600 000 instances. The results have showed that KNIME can handle larger datasets than WEKA, when applying some classifi cation methods. The accuracy of classifi cation is high enough, when the classifi cation methods, implemented in the systems, are used. %; font-family: Calibri, sans-serif;">