Independent Component Analysis for Ensemble Predictors with Small Number of Models

The article presents independent component analysis (ICA) applied to the concept of ensemble predictors. The use of ICA decomposition enables to extract components with particular statistical properties that can be interpreted as destructive or constructive for the prediction. Such process can be treated as noise filtration from multivariate observation data, in which observed data consist prediction results. As a consequence of the ICA multivariate approach, the final results are combination of the primary models, what can be interpreted as aggregation step. The key issue of the presented method is the identification of noise components. For this purpose, a new method for evaluating the randomness of the signals was developed. The experimental results show that presented approach is effective for ensemble prediction taking into account different prediction criteria and even small set of models.