Abstract
INTRODUCTION: Data clustering is an important field of machine learning that has applicability in wide areas, like, business analysis, manufacturing, energy, healthcare, traveling, and logistics. A variety of clustering applications have already been developed. Data clustering approaches based on self-organizing map (SOM) generally use the map dimensions (of the grid) ranging from 2 × 2 to 8 × 8 (4-64 neurons [microclusters]) without any explicit reason for using the particular dimension, and therefore optimized results are not obtained. These algorithms use some secondary approaches to map these microclusters into the lower dimension (actual number of clusters), like, 2, 3, or 4, as the case may be, based on the optimum number of clusters in the specific data set. The secondary approach, observed in most of the works, is not SOM and is an algorithm, like, cut tree or the other. METHODS: In this work, the proposed approach will give an idea of how to select the most optimal higher dimension of SOM for the given data set, and this dimension is again clustered into the lower actual dimension. Primary and secondary, both utilize the SOM to cluster the data and discover that the weight matrix of the SOM is very meaningful. The optimized two-dimensional configuration of SOM is not the same for every data set, and this work also tries to discover this configuration. RESULTS: The adjusted randomized index obtained on the Iris, Wine, Wisconsin diagnostic breast cancer, New Thyroid, Seeds, A1, Imbalance, Dermatology, Ecoli, and Ionosphere is, respectively, 0.7173, 0.9134, 0.7543, 0.8041, 0.7781, 0.8907, 0.8755, 0.7543, 0.5013, and 0.1728, which outperforms all other results available on the web and when no reduction of attributes is done in this work. CONCLUSIONS: It is found that SOM is superior to or on par with other clustering approaches, like, k-means or the other, and could be used successfully to cluster all types of data sets. Ten benchmark data sets from diverse domains like medical, biological, and chemical are tested in this work, including the synthetic data sets.