Abstract

In this paper, the issues of classification based on dispersed data are considered. For this purpose, an approach is used in which prediction vectors are generated locally using the k-nearest neighbors classifier. However, in central server, the final fusion of prediction vectors is made with the use of a neural network. The main aim of the study is to check the influence of various data characteristics (the number of conditional attributes, the number of objects, the number of decision classes) and the degree of dispersion and noise intensity on the quality of classification of the considered approach. For this purpose, 270 data sets were generated that differed by the above factors. Experiments were carried out using these data sets and statistical tests were performed. It was found that each of the examined factors has a statistically significant impact on the quality of classification. However, the number of conditional attributes, degree of dispersion, and noise intensity have the greatest impact. Multidimensionality in dispersed data affects the results positively, but the analyzed method is only resistant to a certain degree of noise intensity and dispersion.

Recommended Citation

Przybyła-Kasperek, M. & Marfo, K. F. (2022). Influence of Noise and Data Characteristics on Classification Quality of Dispersed Data Using Neural Networks on the Fusion of Predictions. In R. A. Buchmann, G. C. Silaghi, D. Bufnea, V. Niculescu, G. Czibula, C. Barry, M. Lang, H. Linger, & C. Schneider (Eds.), Information Systems Development: Artificial Intelligence for Information Systems Development and Operations (ISD2022 Proceedings). Cluj-Napoca, Romania: Babeș-Bolyai University.

Share

COinS
 

Influence of Noise and Data Characteristics on Classification Quality of Dispersed Data Using Neural Networks on the Fusion of Predictions

In this paper, the issues of classification based on dispersed data are considered. For this purpose, an approach is used in which prediction vectors are generated locally using the k-nearest neighbors classifier. However, in central server, the final fusion of prediction vectors is made with the use of a neural network. The main aim of the study is to check the influence of various data characteristics (the number of conditional attributes, the number of objects, the number of decision classes) and the degree of dispersion and noise intensity on the quality of classification of the considered approach. For this purpose, 270 data sets were generated that differed by the above factors. Experiments were carried out using these data sets and statistical tests were performed. It was found that each of the examined factors has a statistically significant impact on the quality of classification. However, the number of conditional attributes, degree of dispersion, and noise intensity have the greatest impact. Multidimensionality in dispersed data affects the results positively, but the analyzed method is only resistant to a certain degree of noise intensity and dispersion.