Information-theoretic selection of high-dimensional spectral features for structural recognition

Pattern recognition methods often deal with thousands of features. Therefore, dimensionality reduction is crucial to make datasets tractable. We propose a novel feature selection method for supervised classification within an information-theoretic framework. Mutual information is exploited for measuring the statistical relation between a subset of features and the class labels of the samples. Traditionally, mutual information has been only measured for ranking single features. We analyze different estimation methods which estimate entropy and mutual information directly from the set of samples, and allow us to efficiently evaluate multivariate sets of thousands of features. We experiment with unattributed spectral graph features extracted from 3D shapes, and show what is the contribution of each spectral feature to graph classification. Besides succeeding in classifying graphs from structure only, we test the robustness of selected spectral features to perturbations of the dataset.