Direct classification of high-dimensional data in low-dimensional projected feature spaces--comparison of several...

Area of Science:

• Computational biology
• Data visualization
• Machine learning

Background:

• High-dimensional data presents challenges in visualization and classification.
• Existing methods may not fully preserve relative relationships between data points.
• The Relative Distance Plane (RDP) was previously introduced for visualization.

Purpose of the Study:

• To extend the applicability of the RDP mapping from data visualization to classification.
• To develop and evaluate new classification methods utilizing the RDP.
• To improve classification accuracy and confidence for high-dimensional biomedical data.

Main Methods:

• Direct application of RDP mapping to standard classifiers like Linear Discriminant Analysis (LDA) and nearest neighbor methods.
• Development of a novel transvariation probabilities-based classification method within the RDP framework.
• Creation of new coordinate systems by combining reference directions for arbitrary classifier development.
• Implementation of a weighted accuracy approach by cycling through reference pairs to enhance classification confidence.

Main Results:

• Demonstrated successful extension of RDP mapping for classification tasks.
• Achieved increased confidence in classification results through weighted accuracy.
• Evaluated classification performance on multiple high-dimensional biomedical datasets.
• Showcased the versatility of RDP in accommodating various classification algorithms.

Conclusions:

• The RDP mapping is a powerful tool for both visualization and classification of high-dimensional data.
• RDP-based classifiers offer enhanced accuracy and reliability, particularly for complex biomedical datasets.
• The RDP framework provides a flexible platform for developing and integrating diverse classification strategies.