Supervised learning in presence of outliers, label noise and unobserved classes

Three important issues are often encountered in Supervised Classifica-\r\ntion: class-memberships are unreliable for some training units (Label Noise), a pro-\r\nportion of observations might depart from the bulk of the data structure (Outliers) and\r\ngroups represented in the test set may have not been encountered earlier in the learn-\r\ning phase (Unobserved Classes). The present work introduces a Robust and Adaptive\r\nEigenvalue-Decomposition Discriminant Analysis (RAEDDA) capable of handling\r\nsituations in which one or more of the afore described problems occur. Transductive\r\nand inductive robust EM-based procedures are proposed for parameter estimation and\r\nexperiments on real data, artificially adulterated, are provided to underline the benefits\r\nof the proposed method.