Mutual information based feature selection for acoustic autism diagnosis

Pervasive Developmental Disorders (PDD) are known to affect children's social interactions and mental development. Prosodic and linguistic cues can be used to diagnose the disorders at early ages. Computational paralinguistics can be applied for tele-monitoring and/or educating the children with PDD. For better understanding the disorders, a small subset of highly informative features is needed. From machine learning perspective, feature selection (FS) is an important step for generalization ability of the learner and drawing inferences about the underlying problems. Since, the high dimensional data are vulnerable to comprise redundant and irrelevant features. The most popular FS methods depend on Mutual Information (MI), that resort to discretization of features. Though the effect of different discretization schemes are studied in literature, to the best of our knowledge the effect of different number of bins for equal width z-score discretization is not studied for MI based FS. Since MI computation depends on the number of discrete categories, we hypothesize that the feature ranking and therefore performance trajectory also changes. We carry out extensive experiments using eight MI based FS methods on the INTERSPEECH 2013 Autism sub-challenge corpus. The comparative results verify our hypothesis and lead to interesting remarks for future studies. Also in this thesis, adjustment for chance factor is proposed for normalizing MI measures, therefore obtaining a new MI based FS criterion. Finally, we choose the candidate ranked features by considering the effect of discretization, and achieve 70.68\% Unweighted Average Recall (UAR) performance on the test set using only 2\% of the feature set. This result advances state-of-the-art performance on the test set adhering to the challenge protocol.