IGSOM: Incremental Clustering Based on Self-Organizing-Mapping

Abstract

Because of today's explosive information from Internet, people will contact much new information at any moment. So how to analyze this non-stationary information becomes more and more important. Clustering analysis is a good information analysis method, but many clustering algorithms only fit to stationary situation. Then in this paper, a novel incremental clustering algorithm based on self-organizing-mapping-IGSOM is provided to dispose this non-stationary information. This algorithm first uses self-organizing-mapping algorithm to construct a neuron model from original data. Then it selects some sample data from this neuron model, and combines the samples with new coming data together to train a new neuron model. To solve unbalance between sample data and new coming data, it alters sample data's weights. The experiments demonstrate that this incremental clustering method can dispose non-stationary data well, and has relatively high precision. Because only small samples are selected to replace large-scale original data, clustering time is also short.