Multiview Feature Selection for Single-View Classification.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Real-world applications often generate multiview data during development.
• Testing with single-view data is preferred due to cost and availability constraints, despite multiview training data.

Purpose of the Study:

• To develop a multiview feature selection method for effective single-view testing.
• To leverage knowledge from all views to guide feature selection within individual views.

Main Methods:

• A multiview feature weighting scheme is proposed.
• This scheme maximizes local margins and preserves inter-view sample similarities.
• The method enables cross-view matching with pre-computed weights.

Main Results:

• The method was evaluated on nine real-world datasets and three biometric recognition applications.
• It demonstrated significant improvements in classification error rates.
• Achieved an average reduction of 31% in classification error rate compared to state-of-the-art methods.

Conclusions:

• The proposed multiview feature selection method effectively utilizes all available views for improved single-view performance.
• The technique offers a practical solution for scenarios where multi-view testing is infeasible.
• The method shows strong potential for applications in biometric recognition and other fields utilizing multiview data.