Orthogonal self-guided similarity preserving projection for classification and clustering

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Traditional dimensionality reduction (DR) methods often rely on original features, which may be noisy or redundant, leading to unreliable intrinsic data structure estimation.
• A key challenge is the interdependence between optimal feature representation and accurate intrinsic data structure.

Purpose of the Study:

• To propose a unified learning framework that simultaneously optimizes feature representation and intrinsic data structure.
• To address the limitations of existing methods in capturing accurate data structure and extracting optimal features.

Main Methods:

• A novel unified learning framework is developed to learn the data structure from feature learning results.
• Features are learned to preserve the refined data structure, leveraging the interaction between feature representation and structure determination.
• The framework simultaneously obtains the optimal feature representation and intrinsic data structure.

Main Results:

• Experimental results show superior performance compared to state-of-the-art methods in dimensionality reduction (DR) and subspace clustering.
• The proposed method effectively captures accurate data structure and obtains optimal feature representations.
• The framework demonstrates the advantage of jointly learning feature representation and intrinsic data structure.

Conclusions:

• The unified learning framework successfully addresses the dilemma of feature representation and intrinsic structure.
• The method offers improved accuracy in dimensionality reduction and subspace clustering tasks.
• The developed approach provides a more robust way to handle noisy or redundant features in data analysis.