Contrastive independent subspace analysis network for multi-view spatial information extraction.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view classification typically uses semantic information, often overlooking crucial spatial data.
• Integrating spatial information enhances data representation and correlation, vital for classification performance.

Purpose of the Study:

• To develop a novel framework for multi-view classification that effectively incorporates spatial information.
• To improve classification accuracy by extracting and utilizing latent spatial features from multi-view data.

Main Methods:

• Proposed a robust independent subspace analysis network (ISA) optimized by sparse and soft orthogonal optimization.
• Developed a contrastive independent subspace analysis (CISA) framework for enhanced spatial perspective optimization.
• Implemented contrastive subspace optimization for separating subspaces and contrastive fusion optimization for cross-view correlations.

Main Results:

• Achieved state-of-the-art accuracies on four benchmark multi-view datasets (76.95%, 98.50%, 93.33%, 88.24%).
• Significantly outperformed the second-best method by up to 8.57% in accuracy.
• Visualization experiments confirmed the effectiveness of subspace and feature space optimization.

Conclusions:

• The proposed MvCISA framework effectively utilizes spatial information for superior multi-view classification.
• The method demonstrates significant performance gains and holds potential for various downstream tasks.
• Spatial feature extraction and fusion are critical for advancing multi-view classification techniques.