A comparative study of similarity measurement in a new neural model of Unsupervised clustering
Hicham Ohmaid, Meriem Timouyas, Souad Eddarouich · 2019
The clustering is one of the most active research and application areas of neural networks. Moreover, it is widely used in many disciplines as data mining, machine learning, decision-making, bioinformatics, image segmentation, and many other areas. The big challenge for clustering data is to choose the right criterion of resemblance for a given data set, which makes distance measure an important factor. In this paper, we present a new statistical neural clustering approach based on Hebbian competitive training and we study three distance measures: Euclidean, Mahalanobis and Manhattan, and their effect on the clustering of different unlabeled data sets forms. Compared to classical approaches such as K-means, this approach has the advantage of high speed because of using a neural network equipped with parallel information processing. Additionally, the proposed procedure does not pass by any thresholding and does not require any information a priori on the number of classes nor on the structure of their distributions in the sample.