Stacking Class Probabilities Obtained from View-Based Cluster Ensembles

Abstract

In pattern recognition applications with high number of input features and insufficient number of samples, the curse of dimensionality can be overcome by extracting features from smaller feature subsets. The domain knowledge, for example, can be used to group some of the features together, which are also known as "views". The features extracted from views can later be combined (i.e. stacking) to train a final classifier. In this work, we demonstrate that even very simple features such as class-distributions within clusters of each view can serve as such valuable features.