Abstract:

In one-class classifier(OCC)  design, considering the structure of the target data is a possible way to improve the generalization ability of the model. However, while the targets follow multicluster distributions, it is more reasonable to consider each cluster’s structure individually rather than just to treat all of them as a whole. The novel algorithm  structure large margin OCC(SLMOCC) fulfills the above strategy by restricting each data’s Mahalanobis distance to the hyperplane. Through maximizing the minimum Mahalanobis margin, SLMOCC is able to find the more reasonable optimal hyperplane attributed to its finer cluster granularity description compared with other alternatives. As for extracting the underlying data structure, this work adopts the Ward’s agglomerative hierarchical clustering on input data or data mapping in kernel space. Experimental results on toy data and UCI benchmark datasets have shown that SLMOCC outperforms  other structural OCCs.
 

Key words: one-class classifier;cluster structure information, Mahalanobis distance