AC-ViSOM: hybridising the modified adaptive coordinate (AC) and ViSOM for data visualization

Abstract

ViSOM's (Visualization induced SOM) final map can be seen as a smooth net embedded in the input space, where the distances among neurons are controlled by a regularization control parameter which is usually heuristically chosen on a trial and error basis. Empirical studies shown that ViSOM suffers from dead neuron problem, since a big number of neurons fall outside of the data region due to the regularization effect, even though regularization control parameter is properly chosen. In this paper, a modified Adaptive Coordinate (AC) approach that is able to preserve data structure is hybridised with ViSOM is proposed. Experimental studies on benchmark datasets shown that the proposed method was able to eliminate the selection of regularization control parameter and minimizing the dead neuron for better data representation.