Scalable Multi-View Semi-Supervised Classification via Adaptive Regression.

Area of Science:

• Machine Learning
• Image Processing
• Computer Vision

Background:

• Multi-view learning is crucial with the rise of multi-view data.
• Labeled data scarcity poses challenges in many machine learning applications.
• Semi-supervised classification is vital for leveraging unlabeled data.

Purpose of the Study:

• To propose an effective algorithm for multi-view semi-supervised classification.
• To address the challenge of limited labeled data in multi-view learning.
• To develop a method robust to low-quality data views.

Main Methods:

• Developed the Multi-View Semi-Supervised Classification via Adaptive Regression (MVAR) algorithm.
• Utilized regression-based loss functions with L2,1 matrix norm for each view.
• Formulated the objective function as a weighted combination of individual view losses.
• Designed an efficient algorithm with proven convergence for non-smooth L2,1-norm minimization.
• Incorporated adaptive weight coefficients to balance view contributions.

Main Results:

• The proposed MVAR algorithm demonstrated efficiency in calculations, suitable for large-scale datasets.
• Adaptive weights improved robustness against low-quality views.
• Learned projection matrices and bias vectors facilitate out-of-sample data prediction.
• Experimental validation on real-world datasets and scene classification confirmed MVAR's effectiveness compared to benchmark methods.

Conclusions:

• MVAR is an effective and efficient algorithm for multi-view semi-supervised classification.
• The adaptive regression approach enhances robustness and scalability.
• The method shows strong performance in practical applications like scene classification.