Path for Kernel Adaptive One-Class Support Vector Machine

Van-Khoa Le (1) , Pierre Beauseroy (1)

1 LM2S - Laboratoire Modélisation et Sûreté des Systèmes

This paper proposes a Kernel Adaptive One Class SVM (KAOC-SVM) method based on the model introduced by A. Scholkopf and al. The aim is to find the solution path - the path of Lagrange multiplier a - as the kernel parameter changes from one value to another. It is similar to the regularization path approach proposed by Hastie and al., which finds the path when the regularization parameter ? changes from 0 to 1. In present case, the main difference is that the Lagrange multiplier paths are not piecewise linear anymore. Experimental results show that the proposed method is able to compute one-class SVMs with the same accuracy as traditional method but exploring all solutions combining 2 kernels. Simulation results are presented and CPU requirement is analyzed.

Domains: Machine Learning [cs.LG]

Van-Khoa Le, Pierre Beauseroy. Path for Kernel Adaptive One-Class Support Vector Machine. 2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA), Dec 2015, Miami, United States. pp.503-508, ⟨10.1109/ICMLA.2015.127⟩. ⟨hal-02358669⟩