Large margin distribution multi-class supervised novelty detection.

As one of state-of-the-art supervised novelty detection models, support vector machine-supervised novelty detection (SVM-SND) can recognize whether a test instance is a novelty or which class the test instance comes from if it is a normal instance through a single model. The SVM-SND inherits the idea of maximizing minimum margin from one-class support vector machine. However, current research has proved that maximizing minimum margin is not the necessary condition to ensure the generalization in classification. This paper introduces margin distribution into multi-class supervised novelty detection and proposes a large margin distribution-supervised novelty detection (LMD-SND) model in which margin distribution is used to enhance the performance for multi-class supervised novelty detection. The margin distribution is represented by margin mean and margin variance. The LMD-SND maximizes the sum of margin means of all classes and minimizes the sum of margin variance of all classes through extra two terms, simultaneously. Compared with SVM-SND, LMD-SND inherits its merits for multi-class supervised novelty detection but has better generalization. The experimental results show that LMD-SND is superior to SVM-SND and can demonstrate comparative performance with shallow multi-class supervised novelty detection models, KNFST and KLMNN, and deep novelty detection models, DSVDD and DMSVDD.