Search

Novelty detection involves modeling the normal patterns for detecting any divergence from this behavior. Our recently proposed algorithm, Glabal&Local One Class Classifier (GLocal OCC), can solve this problem by maximizing the margin between the hyperplane and the origin through embedding the global information into the OCSVM framework. In this paper, we propose Linear Programming (LP) GLocal OCC (lpGLocal OCC) instead of the Quadratic Programming optimization to speed up GLocal OCC. By minimizing the average functional distance of the overall samples to the hyperplane, the lpGLocal OCC can attract the optimal hyperplane towards the centre of the data without using the origin anymore. Borrow off-the-shelf LP solver, this novel algorithm can be implemented easily and process solve large datasets rapidly. Results on benchmark datasets show that lpGLocal OCC not only has the comparable generalization power compared with the GLocal OCC besides its efficiency, but also has better generalization than (Ip)OCSVM due to its structured learning approach. [ABSTRACT FROM AUTHOR]