Collective Affinity Learning for Partial Cross-Modal Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Unsupervised hashing methods are crucial for cross-modal retrieval.
• Real-world data often presents incomplete modalities with missing samples.
• Existing methods struggle with partial multi-modal data due to assumptions of complete data presence.

Purpose of the Study:

• To propose a novel method, Collective Affinity Learning Method (CALM), for cross-modal retrieval with partial multi-modal data.
• To address the limitations of existing methods that assume complete data presence.
• To develop a robust approach for generating binary codes from incomplete multi-modal datasets.

Main Methods:

• CALM collectively constructs modality-specific bipartite graphs.
• A probabilistic model derives complete data-to-anchor affinities, recovering missing information.
• A robust model fuses affinities by adaptively learning a unified anchor graph.
• Neighborhood information from the anchor graph provides feedback for affinity reconstruction.
• Anchor Graph Hashing (AGH) is applied for cross-modal retrieval.

Main Results:

• Theoretical analysis confirms the method's ability to recover missing adjacency information.
• The developed algorithm exhibits linear time complexity and fast convergence.
• Experimental results on benchmark datasets demonstrate CALM's superior performance over existing methods.
• CALM consistently outperforms current approaches in cross-modal retrieval tasks with incomplete data.

Conclusions:

• CALM effectively handles partial multi-modal data by adaptively learning an anchor graph.
• The proposed method offers a robust solution for cross-modal retrieval challenges posed by missing data.
• CALM provides a significant advancement in unsupervised hashing for incomplete multi-modal datasets.