Supervised Variational Relevance Learning, An Analytic Geometric Feature Selection with Applications to Omic...

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Distance-based similarity is crucial for pattern classification.
• Existing methods may not optimally define class separability.
• Relevance learning offers a framework for feature relevance determination.

Purpose of the Study:

• To introduce Supervised Variational Relevance Learning (Suvrel), a variational method for metric tensor determination.
• To enhance distance-based similarity in pattern classification.
• To improve classification efficiency and accuracy.

Main Methods:

• Developed a variational method inspired by relevance learning.
• Applied the method to a cost function penalizing intraclass distances and favoring interclass distances.
• Analytically derived the metric tensor minimizing the cost function.
• Utilized linear transformations with the metric tensor for preprocessing patterns.

Main Results:

• The metric tensor effectively preprocesses data for more efficient classification.
• Tested on publicly available datasets with standard classifiers.
• Achieved improved performance compared to MAQC-II project results on relevant datasets, even without additional preprocessing.

Conclusions:

• Suvrel provides an effective approach for learning metric tensors in pattern classification.
• The method enhances distance-based similarity, leading to improved classification performance.
• Suvrel demonstrates potential for advancing pattern recognition tasks.