A Local-Density-Ratio Based Algorithm for Setting Weight in Weighted Least Squares Support Vector Machine
Zhuangfeng Shao, Xiaowei Yang, Wen Wen, Zhifeng Hao · 2006
How to assign weights on samples is an important subject in weighted least squares support vector machine (WLS-SVM) for regression problems, which largely influences the robustness of the WLS-SVM. Based on the local outlier factor (LOF), a useful factor for detecting outlier in knowledge discovery, we propose a local-density-ratio (LDR) based weight-setting algorithm for WLS-SVR in this paper. In the proposed algorithm, weights are assigned to the samples according to their neighborhood density ratios. In order to simplify the parameter selection, a single parameter strategy is introduced, which avoids choosing two thresholds in other heuristic weight-setting strategies. Numerical experiments show that the proposed algorithm is able to distinguish most of the noises and produces robust estimator