Spectral Clustering and Long Timeseries Classification

Abstract Clustering is a method of allocating data points in various groups, known as clusters, based on similarity. The notion of expressing similarity mathematically and then maximizing it (minimize dissimilarity) can be formulated as an optimization problem. Spectral clustering is an example of such an approach to clustering, and it has been successfully applied to visualization of clustering and mapping of points into clusters in two and three dimensions. Higher dimension problems remained untouched due to complexity and, most importantly, lack of understanding what “similarity” means in higher dimensions. In this paper, we apply spectral clustering to long timeseries EEG (electroencephalogram) data. We developed several models, based on different similarity functions and different approaches for spectral clustering itself. The results of the numerical experiment demonstrate that the created models are accurate and can be used for timeseries classification.