A class of single-class minimax probability machines for novelty detection

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Single-class minimax probability machines (MPMs) provide robust novelty detection with distribution-free worst-case bounds.
• Traditional MPMs exhibit a high false negative rate (FNR) due to their conservative nature in identifying outliers.

Purpose of the Study:

• To develop a more aggressive single-class MPM to address the high FNR of existing methods.
• To bound the best-case probability of a pattern falling within the normal region for improved outlier detection.

Main Methods:

• Proposed a novel, more aggressive single-class MPM.
• Utilized a pair of MPMs (traditional and aggressive) to define the solution space.
• Employed a middle hyperplane between the two MPMs for optimal classification.

Main Results:

• The proposed aggressive MPM effectively reduces the false negative rate.
• The combined approach achieves a better balance between false positives (FPs) and FNs.
• Experiments on real-world datasets demonstrate encouraging improvements in novelty detection accuracy.

Conclusions:

• The enhanced single-class MPM offers a superior compromise between precision and recall.
• This method provides a more effective solution for novelty detection tasks with high-stakes applications.
• The dual-MPM approach refines outlier identification by carefully delimiting the normal region.