Optimizing Kohonen Classification of Mixed Data with Partial Distance and Referent Vector Initialization

Abstract

The success of neural network models in clustering problems is highly dependent on the quality and diversity of the data used. Self-organizing maps (SOM), a semi-supervised data learning tool introduced by Kohonen in the 1980s, have been widely used in various fields such as signal and text recognition, industrial data analysis, speech and image recognition, etc. SOM's competitive learning clustering method, where each node specializes in a specific subset of data, has proven to be a powerful technique. In this paper, we propose a new SOM variant suitable for handling numerical, interval, and categorical attributes simultaneously. Instead of random initialization of weights, we utilize the ASAICC algorithm to select initial referent vectors.  Furthermore, we suggest representing one cluster using multiple referent vectors at once. The effectiveness of the proposed Kohonen variant is evaluated using well-known benchmark datasets, and the results are reported using reliable performance metrics. The simulation of the new algorithm is conducted using the R language, and the obtained results demonstrate the superiority of the proposed approach.