Nonlinear Prediction Based on Independent Component Analysis Mixture Modelling

This paper presents a new algorithm for nonlinear prediction based on independent component analysis mixture modelling (ICAMM). The data are considered from several mutually-exclusive classes which are generated by different ICA models. This strategy allows linear local projections that can be adapted to partial segments of a data set while maintaining generalization (capability for nonlinear modelling) given the mixture of several ICAs. The resulting algorithm is a general purpose technique that could be applied to time series prediction, to recover missing data in images, etc. The performance of the proposed method is demonstrated by simulations in comparison with several classical linear and nonlinear methods.