Dynamic clustering of time series with Echo State Networks

Abstract

In this paper we introduce a novel methodology for unsupervised analysis of time series, based upon the iterative implementation of a clustering algorithm embedded into the evolution of a recurrent Echo State Network. The main features of the temporal data are captured by the dynamical evolution of the network states, which are then subject to a clustering procedure. We apply the proposed algorithm to time series coming from records of eye movements, called saccades, which are recorded for diagnosis of a neurodegenerative form of ataxia. This is a hard classification problem, since saccades from patients at an early stage of the disease are practically indistinguishable from those coming from healthy subjects. The unsupervised clustering algorithm implanted within the recurrent network produces more compact clusters, compared to conventional clustering of static data, and provides a source of information that could aid diagnosis and assessment of the disease.