Maximum Decentral Projection Margin Classifier for High Dimension and Low Sample Size problems.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Manual data labeling is labor-intensive, even with automation.
• High Dimension and Low Sample Size (HDLSS) data present challenges in machine learning due to data piling and approximate equidistance.
• Traditional classifiers often perform poorly on HDLSS data classification problems.

Purpose of the Study:

• To propose a novel classifier, the Maximum Decentral Projection Margin Classifier (MDPMC), for handling HDLSS data.
• To integrate constraints for maximizing projection distance into a Support Vector Classifier (SVC) framework.
• To address the data piling and approximate equidistance issues inherent in HDLSS datasets.

Main Methods:

• Developed the Maximum Decentral Projection Margin Classifier (MDPMC) within the Support Vector Classifier (SVC) framework.
• Incorporated maximization of projection distance between decentralized points and supporting hyperplanes.
• Evaluated MDPMC on ten real-world HDLSS datasets.

Main Results:

• MDPMC effectively addresses data piling and approximate equidistance problems in high-dimensional spaces.
• Experimental results demonstrate MDPMC's superior performance on HDLSS data.
• MDPMC achieved better predictive accuracy and lower classification errors compared to seven other classifiers, including SVC variants, DWD, and PGLMC.

Conclusions:

• The proposed MDPMC classifier offers a robust solution for classification tasks involving High Dimension and Low Sample Size data.
• MDPMC enhances predictive accuracy and reduces classification errors on challenging HDLSS datasets.
• This approach provides a significant improvement over existing methods for HDLSS data analysis.