Ranking Principal Components in Face Spaces through AdaBoost.M2 Linear Ensemble
Tiene A. Filisbino, Gilson Antonio Giraldi, Carlos Eduardo Thomaz · 2016
Despite the success of Principal Component Analysis (PCA) for dimensionality reduction, it is known that its most expressive components do not necessarily represent important discriminant features for pattern recognition. In this paper, the problem of ranking PCA components, computed from multi-class databases, is addressed by building multiple linear learners that are combined through the AdaBoost.M2 in order to determine the discriminant contribution of each PCA feature. In our implementation, each learner is a weakened version of a linear support vector machine (SVM). The strong learner built by the ensemble technique is processed following a strategy to get the global discriminant vector to sort PCA components according to their relevance for classification tasks. Also, we show how the proposed methodology to compute the global discriminant vector can be applied to other multi-class approaches, like the linear discriminant analysis (LDA). In the computational experiments we compare the obtained approaches with counterpart ones using facial expression experiments. Our experimental results have shown that the principal components selected by the proposed technique allows higher recognition rates using less linear features.