Abstract:

To improve the generalization ability of one-class classifier, more prior knowledge were taken into account on the existed models. A new structured one-class support vector machine with local density embedding (ldSOCSVM) was proposed, which could embed local information of target data into the structured one-class support vector machine (SOCSVM). By means of K-nearest neighbor, the weighted factor was extracted and applied to the corresponding samples by fully utilizing local information with the global ones inherited from SOCSVM, the ldSOCSVM improved the generalization ability. Experimental results on UCI datasets showed that the proposed classifier could achieve better generalization capability compared with related algorithms.

Key words: one-class classifier, prior knowledge, structured one-class support vector machine, local density, weighted factor