Efficient Unsupervised Parameter Estimation for One-Class Support Vector Machines.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• One-class support vector machines (OCSVMs) are effective for anomaly detection but sensitive to hyperparameter settings.
• Real-world data often lacks clean training sets, necessitating unsupervised or semi-supervised OCSVM application.
• Anomalies in training data can skew OCSVM boundaries, reducing detection accuracy.

Purpose of the Study:

• To propose a novel technique for setting OCSVM hyperparameters and cleaning unlabeled training sets.
• To improve the performance and efficiency of OCSVMs in anomaly detection tasks.
• To address the challenges posed by noisy training data in OCSVM applications.

Main Methods:

• A new technique combining K-nearest neighbors for anomaly removal and direct hyperparameter estimation.
• Evaluation on diverse benchmark datasets with varying distributions and dimensionality.
• Comparison against supervised, semi-supervised, and unsupervised hyperparameter estimation methods.

Main Results:

• The proposed method is significantly faster (up to 70x) than supervised grid-search and cross-validation.
• It achieves speed improvements of one to three orders of magnitude over existing semi-supervised and unsupervised methods.
• The technique demonstrates statistically superior performance and comparable accuracy to supervised methods.

Conclusions:

• The proposed method offers an efficient and effective solution for hyperparameter tuning and data cleaning in OCSVMs.
• It overcomes limitations of traditional unsupervised and semi-supervised approaches for anomaly detection.
• This technique enhances the practical applicability of OCSVMs in real-world scenarios with imperfect data.