MST-GEN: An Efficient Parameter Selection Method for One-Class Extreme Learning Machine.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• One-class classification (OCC) models target data from a single class to identify outliers.
• Existing OCC methods like one-class support vector machine (OCSVM) and support vector data description (SVDD) are widely used.
• One-class extreme learning machine (OCELM) offers fast learning and competitive performance but suffers from difficult parameter selection.

Purpose of the Study:

• To propose an automatic parameter selection method, MST-GEN, for OCELM.
• To address the under-explored challenge of parameter optimization in OCELM and similar OCC models.
• To enhance the efficiency and accuracy of OCC model performance.

Main Methods:

• Constructing an n-round minimal spanning tree (MST) to represent target data structure and distribution.
• Generating pseudo-outliers using edge pattern detection and a novel 'repelling' process based on MST information.
• Generating pseudo-target data utilizing the MST to bypass time-consuming cross-validation for accelerated parameter selection.

Main Results:

• MST-GEN demonstrates highly efficient and accurate parameter selection for OCELM compared to existing methods.
• The proposed method enables OCELM to achieve improved performance in various OCC applications.
• MST-GEN shows favorable applicability to other prevalent OCC methods, including OCSVM and SVDD.

Conclusions:

• MST-GEN provides an effective solution for the parameter selection problem in OCELM.
• The method significantly enhances the practical utility and performance of OCELM and other OCC algorithms.
• This approach offers a more efficient and accurate alternative for optimizing OCC models.