Kernel Choice in One-Class Support Vector Machines for Novelty Detection
Qinong Tian, Peixi Liu, Tian Yi Wu, Ziyue Chen · Atlantis Highlights in Computer Sciences/Atlantis highlights in computer sciences · 2023
This work concentrates on novelty detection, a semi-supervised learning problem concerned with deciding if the new observation is sufficiently different from the ones seen so far.This paper mainly considers a variant of the support vector classification approach, which estimates the contours of the distribution of the initial observations and then can be used to decide if the new observations are abnormal.We try to estimate a negative function on the outlier points in the input space and a positive on the complement.A kernel expansion gives this decision function.The effectiveness of this kernel method is closely related to the choice of kernel functions and hyperparameters.Due to the demand for general and effective hyperparameter selection regulations, we investigate three approaches, including GridSearch in Python, median heuristic and Bayesian kernel learning.Some relevant experiments are performed in this paper.According to the experiments, we have learned that the choice of kernels and parameters can greatly influence the detection result.