Deep Multi-View Clustering With Meta Information Compression.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multi-view clustering methods utilize data consistency and complementarity for sample partitioning.
• Existing deep learning approaches struggle to balance complementary information with essential details, leading to redundancy or information loss.

Purpose of the Study:

• To develop a novel meta-learning based method for learning clustering-friendly representations with minimal redundancy.
• To address the limitations of existing deep learning methods in multi-view clustering.

Main Methods:

• A meta-learning framework employing bi-level optimization for feature embedding and an information compressor.
• Training an information compressor to create compact representations with minimized redundancy.
• Implementing a semantic puzzle mechanism to integrate semantic fragments and enhance discriminative power.

Main Results:

• The proposed method effectively learns key semantics with reduced redundancy.
• The semantic puzzle mechanism successfully complements semantic fragments, creating a consensus representation.
• Extensive experiments demonstrated significant performance improvements over state-of-the-art methods on various datasets.

Conclusions:

• The novel meta-learning approach offers a superior strategy for multi-view clustering.
• The method effectively overcomes the dilemma of information redundancy versus loss in deep clustering.
• The learned representations exhibit strong discriminative power, leading to enhanced clustering accuracy.