Logistic Discrimination with Total Variation Regularization

This article introduces a regularized logistic discrimination method that is especially suited for discretized stochastic processes (such as periodograms, spectrograms, EEG curves, etc.). The proposed method penalizes the total variation of the discriminant directions, giving smaller misclassification errors than alternative methods, and smoother and more easily interpretable discriminant directions. The properties of the new method are studied by simulation and by a real-data example involving classification of phonemes. Supported in part by NSF Grant DMS-06-04396. Key words and phrases: Classification, discrimination, machine learning, speech recognition. 1 

• Type: Journal
• Year: 2008
• Where: CSSC
• Authors: Robin Rühlicke, Daniel Gervini
• Keywords: CSSC 2008, Discriminant Directions, Interpretable Discriminant Directions, Regularized Logistic Discrimination