Adaptive Feature Projection With Distribution Alignment for Deep Incomplete Multi-View Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Incomplete multi-view clustering (IMVC) faces challenges with inaccurate imputed data and feature distribution discrepancies.
• Existing methods often rely on imputation, which can be unreliable with unknown labels and ignore differences between complete and incomplete data.

Purpose of the Study:

• To propose an imputation-free deep IMVC method that addresses limitations of current approaches.
• To develop a method that learns features while aligning distribution discrepancies in incomplete multi-view data.

Main Methods:

• Utilizes autoencoders for view-specific feature learning and adaptive feature projection to bypass imputation.
• Projects available data into a common feature space to explore cluster information via mutual information maximization.
• Employs a novel mean discrepancy loss, adapted for mini-batch optimization, to achieve distribution alignment.

Main Results:

• The proposed imputation-free method demonstrates comparable or superior performance against state-of-the-art IMVC techniques.
• Experimental results validate the effectiveness of the distribution alignment and adaptive feature projection strategies.
• The method successfully handles missing data without explicit imputation, improving clustering accuracy.

Conclusions:

• The developed imputation-free deep IMVC method offers a robust solution for handling missing data in multi-view clustering.
• Aligning feature distributions and avoiding imputation are key to improving IMVC performance.
• This approach provides a valuable advancement for analyzing incomplete multi-view datasets.