Clustering of objects with poorly formalizable features based on a neural network in the form of Kohonen layers

• Main Authors: I. A. Kubasov, A. V. Melnikov, S. A. Maltsev, I. R. Narushev
• Format: Article
• Language: Russian
• Published: Voronezh state university of engineering technologies 2018-12-01
• Series: Vestnik Voronežskogo Gosudarstvennogo Universiteta Inženernyh Tehnologij
• Subjects: clustering, neural networks, multicriteria analysis
• Online Access: https://www.vestnik-vsuet.ru/vguit/article/view/1811
• ISSN: 2226-910X; 2310-1202

Analysis of profiles of minors in social networks shows that teenagers indicate in them information that increases the level of their social desirability. Such information often does not correspond to the real behavior of the teenager. For a full analysis of the level of deviance of a minor need tools covering the full range of indicators. In contrast to the usual approach to clustering objects based on their Association in groups by the criterion of the minimum distance in multidimensional space when clustering features it is advisable to take into account their proximity to the methods of obtaining information and methods of processing of this information by the inspector for minors. In the first phase of the study is the clustering of signs of deviation, the second the determination of the weighting factors of indicator of the degree of deviance within each group of signs, the third uses the method of cluster-hierarchical approach to forming integral indicator of assessment of deviant behavior of minors. The indicator has a considerable flexibility of the correlation between groups of symptoms and partial characteristics through the introduction of appropriate sets of weighting coefficients. The conclusion is made about the preference of methods based on clustering of objects in the two-dimensional space of targets or accounts of the principal components method, as well as the need for additional analysis of the graphical picture of the relative location of objects. From the comparison of different approaches: 1) clustering on the basis of the generalized indicator of quality and the sign of reverse deviance, 2) clustering on two accounts of the principal components method; 3) clustering on all signs of examination, the following conclusions can be drawn. All methods properly allocate the objects to clusters. However, when you save the main totals (highlighting the best and worst features), the results are slightly different. This is due to the different volume and forms of presentation of the source information. The program assigns numbers of active neurons (clusters) arbitrarily, so in order to arrange the cluster numbers by some feature (for example, the quality of objects), you need to use additional graphical information. From a practical point of view, the first two methods are preferred, based on clustering objects in two-dimensional space, the method of principal components and the analysis of the graphical picture of the mutual location of objects.