POSITION AND DENSITY BASED KERNEL PARAMETER SELECTION FOR ONE-CLASS SVM

One-Class Support Vector Machine (OCSVM) is widely used for anomaly detection, yet its performance is highly sensitive to kernel parameter selection, particularly in the Gaussian kernel. Existing geometry-based criteria rely on extreme-distance measurements, making them highly sensitive to sample sparsity and dimensionality. In this paper, a robust parameter selection framework is proposed for OCSVM that integrates both positional and density information of training samples through a continuous weighting scheme. A statistically motivated objective function is further introduced and incorporates a penalty term to mitigate overfitting. Experiments on benchmark datasets demonstrate that the proposed method consistently achieves performance comparable to or better than two OCSVM settings, Isolation Forest and Local Outlier Factor. The improvements are validated by statistical significance tests, indicating that the proposed method provides a stable and effective solution for kernel parameter selection in OCSVM.