Learning Long-Term Time Series with Generative Topographic Mapping

Abstract

We propose a generative topographic mapping (GTM) based nonlinear model for long-term time series prediction. As a modification of Kohonen self-organizing maps (SOM), GTM has been applied to data classification, visualization and other machine learning problems, however, limited research have been proposed in time series analysis. With a double application of GTM algorithm, a specially designed approach can quantize input data to store temporal evolvement information for trend prediction. Experimental results demonstrate the improved forecast accuracy in long-term trend learning.