Semantic Invariant Multi-View Clustering With Fully Incomplete Information.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Real-world multi-view learning often suffers from incomplete information, such as missing correspondences and instances.
• Current methods typically require paired samples for data correction, which is often impractical due to data collection complexities.

Purpose of the Study:

• To develop a novel framework for multi-view clustering that addresses incomplete information without relying on paired samples.
• To leverage invariant semantic distributions across views to learn consensus semantics and improve clustering performance.

Main Methods:

• The proposed SeMantic Invariance LEarning (SMILE) framework identifies invariant semantic distributions across different views.
• SMILE alleviates cross-view discrepancies to learn consensus semantics, which are robust to distribution shifts.
• These consensus semantics are then used for instance realignment/imputation and cluster formation.

Main Results:

• SMILE demonstrates superior performance compared to 13 state-of-the-art baselines across five benchmark datasets.
• On the NoisyMNIST dataset, SMILE significantly improved clustering accuracy from 19.3%/23.2% to 82.7%/69.0% under full incompleteness of correspondences/instances.
• The framework effectively handles both incomplete correspondences and incomplete instances without paired data.

Conclusions:

• SMILE provides a robust solution for multi-view clustering with incomplete information, eliminating the need for paired samples.
• The method's ability to learn invariant semantic distributions offers a powerful way to handle cross-view discrepancies and data defects.
• The significant performance gains highlight SMILE's potential for practical applications in complex real-world scenarios.