SMILE: Semi-supervised multi-view classification based on dynamical fusion.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised multi-view classification is vital for leveraging complex datasets in fields like medical diagnosis and autonomous driving.
• Existing methods often fail to improve performance due to simple feature fusion and lack of view quality assessment.
• Redundant features and low-quality views hinder the effectiveness of conventional approaches.

Purpose of the Study:

• To propose a novel dynamic fusion approach, SMILE, for enhanced semi-supervised multi-view classification.
• To address the limitations of conventional methods in feature fusion and view quality handling.
• To improve classification accuracy and robustness in multi-view learning scenarios.

Main Methods:

• Developed a high-level semantic mapping module for extracting discriminative features and reducing redundancy.
• Implemented a dynamic fusion module to assess and adaptively weigh the quality of different views per sample.
• Evaluated the SMILE approach against six competitive methods on four diverse datasets.

Main Results:

• SMILE demonstrated significant performance improvements across various evaluation metrics compared to existing methods.
• The dynamic fusion approach effectively mitigated the negative impact of low-quality views on classification.
• Visualization experiments confirmed the method's ability to learn classification-friendly data representations.

Conclusions:

• The proposed SMILE method offers a superior approach to semi-supervised multi-view classification.
• Dynamic feature fusion and quality assessment are key to improving performance in complex multi-view datasets.
• SMILE provides a robust framework for applications requiring accurate multi-view data analysis.