Our paper “Large-Dimensional Characterization of Robust Linear Discriminant Analysis” with Nicolas Auguin and Matthew McKay (HKUST) has just appeared in IEEE Transactions on Signal Processing.
The work focuses on a robust version of linear discriminant analysis (LDA) classifiers, extensively used in machine learning. To account for potential spurious (e.g., non-Gaussian) or mislabeled observations in the training data (i.e., outlying samples), we here propose and characterise the large-dimensional performance (when both the numbers of samples and variables are large) of LDA classifiers employing robust regularised M-estimators of the covariance matrix, essential to the design of a robust discrimination rule when the dataset is corrupted with outliers.