Dimensionality Reduction Based Similarity Visualization for Neural Gas

Abstract

Two commonly used neural networks for vector quantization based analysis of high-dimensional large datasets are the self-organizing map (SOM) and neural gas (NG). Owing to their rigid grid structure, SOMs are widely used for data visualization, whereas NG based visualization has been limited, despite the fact that NG can achieve better quantization than SOMs in terms of quantization error. As a visualization tool for NG, we propose to use a recent projection technique t-SNE (which depends on stochastic neighbor embedding using student t-distribution). T-SNE projection of NG will construct a low-dimensional space where local similarities of high-dimensional data space are preserved to a great extent. In addition, this enables the use of Connives (a topology-based visualization for SOMs) to represent the data space similarities on the low-dimensional projection space. Experiments on the synthetic and real datasets show that the proposed NG visualization based on t-SNE and enhanced with Connives is helpful for interactive analysis of high-dimensional large datasets.