Making Standard SOM Invariant to the Initial Conditions

Abstract

In data clustering, the assessment of learning properties with respect to data is important for a reliable classification. However, in standard Self Organizing Map (SOM), weight vectors initialization is done randomly, leading to a different final feature map each time the initial conditions are changed. To cope with this issue, in this paper, we present a behavioral study of the first iterations of the learning process in standard SOM. After establishing the mathematical foundations of the first passage of input vectors, we show how to conclude a better initialization relatively to the data set, leading to the generation of a unique feature map.