Graph-Collaborated Auto-Encoder Hashing for Multiview Binary Clustering.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Unsupervised hashing methods reduce storage and computation for large-scale data using binary codes.
• Existing methods often overlook local geometric structures and multi-source data complementarity.
• Auto-encoder-based hashing minimizes reconstruction loss, neglecting data consistency.

Purpose of the Study:

• To propose a novel unsupervised hashing algorithm for multiview binary clustering.
• To address limitations of existing methods by incorporating local geometric structure and multi-source data.
• To develop a unified binary code learning approach for enhanced clustering.

Main Methods:

• Developed a Graph-Collaborated Auto-Encoder (GCAE) hashing algorithm.
• Dynamically learned multiview affinity graphs with low-rank constraints.
• Employed collaborative learning between auto-encoders and affinity graphs.
• Incorporated decorrelation and code balance constraints to minimize quantization errors.
• Utilized an alternating iterative optimization scheme for multiview clustering.

Main Results:

• The proposed GCAE hashing method effectively mines underlying geometric information from multiview data.
• Learned a unified binary code by collaborating multiple affinity graphs via an encoder-decoder paradigm.
• Achieved superior performance compared to state-of-the-art alternatives on five public datasets.
• Demonstrated the effectiveness of integrating affinity graphs and auto-encoders for multiview clustering.

Conclusions:

• GCAE hashing provides an effective solution for multiview binary clustering.
• The method successfully captures local geometric structures and leverages multi-source data complementarity.
• The proposed approach offers significant improvements in large-scale data analysis and clustering tasks.