Distance Metrics for Time-Series Data with Concentric Multi-Sphere Self Organizing Maps

Self-Organizing Maps have been shown to be a powerful unsupervised learning a tool in the analysis of complex high dimensional data. SOMs are capable of performing topological mapping, clustering and dimensionality reduction in order to effectively visualize and understand data and it is desirable to apply these techniques to time---series data. In this project a novel approach to time-series learning using Concentric Multi-Sphere SOMs has been expanded and generalized into a unified framework in order to thoroughly test the learning capabilities. It is found that Quantization and Topological Error are not suitable to test the learning performance of the algorithms and it is suggested that future work focus on developing new error measures and learning algorithms.