A non-Gaussian regression algorithm based on mutual information maximization

This article introduces a latent variable regression technique for non-Gaussian distributed variable sets. For a single response variable, a mutual information criterion is blended into the formulation of independent components. Extending this conceptual algorithm to multiple response variables, it reduces to canonical correlation regression if the predictor and response sets are Gaussian distributed. An analysis of the weighted objective function yields that the new algorithm can be reduced to recently published independent component regression methods. Application studies to a simulation example and recorded data confirm that the proposed algorithm can balance between the extraction of latent non-Gaussian components and the accuracy of the regression model.

► We introduce a new latent variable regression technique for non-Gaussian variable sets.
► The new technique is a non-Gaussian extension of Canonical Correlation Regression.
► We contrast the new technique with other regression techniques using 4 case studies.
► The new technique balances between prediction accuracy and source signals extraction.
► Existing regression techniques do not offer the same flexibility and/or accuracy.