Abstract
Ensemble feature selection is known for its robustness and generalization of highly accurate predictive models. In this paper, we use different filter-based feature selection methods in an ensemble manner to improve face recognition. The goal is to distinguish human faces from avatar faces. Our approach was able to achieve very high accuracy, 99%, using less than 1% of the pixels in each image. This was obtained after removing irrelevant features which is known to degrade learning performance and model stability.