Two novel methods for the determination of the number of components in independent components analysis models

Independent Components Analysis is a Blind Source Separation method that aims to find the pure source signals mixed together in unknown proportions in the observed signals under study. It does this by searching for factors which are mutually statistically independent. It can thus be classified among the latent-variable based methods. Like other methods based on latent variables, a careful investigation has to be carried out to find out which factors are significant and which are not. Therefore, it is important to dispose of a validation procedure to decide on the optimal number of independent components to include in the final model. This can be made complicated by the fact that two consecutive models may differ in the order and signs of similarly-indexed ICs. As well, the structure of the extracted sources can change as a function of the number of factors calculated. Two methods for determining the optimal number of ICs are proposed in this article and applied to simulated and real datasets to demonstrate their performance.

► We propose two methods to validate ICA models. ► The first method is based on the correlations between the ICs extracted from several blocks of data. ► The second method is based on the Durbin–Watson criterion. ► Both methods are applied on simulated and real data, and compared to other existing methods. ► They are shown to perform well.