A Supervised Method to Enhance Distance-based Neural Networks' Clustering Performance by Discovering Perfect Representative Neurons

Distance-based neural network clustering has an intrinsic assumption that a particular neuron in the network represents a cluster centroid. However, not all these neurons can perfectly represent the training data; these neurons can only represent part of the training samples. This paper proposes an effective training data splitting method (TDSM) to find perfect representative neurons and improve the clustering results in a distance-based neutral network without changing the original network's internal algorithm or the training data quality. The method allows a network with N neurons to be enlarged to a new m×N neurons network. These neurons represent m sub-networks, and each sub-network perfectly represents a part of the training set with the clustering qualification indicators (purity, Normalized Mutual Information, and Adjusted Rand Index), all equalling 1. Results are statistically validated with a t-test, and we demonstrate that TDSM performs better than the original clustering paradigm on some real data sets.