IB2MC: Information Bottleneck Inspired Balanced Multiview Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Information Theory

Background:

• Unsupervised clustering methods face challenges with information bottleneck (IB) theory due to a lack of prior knowledge and poor clustering security.
• Existing approaches often suffer from performance degradation caused by empty clusters resulting from excessive bias.

Purpose of the Study:

• To propose a scalable framework, Information Bottleneck Inspired Balanced Multiview Clustering (IB2MC), addressing clustering security, optimization synergy, and tuning simplicity.
• To leverage unsupervised IB theory to enhance cluster distribution uncertainty and descriptor discriminability for improved label inference.

Main Methods:

• Developed IB2MC framework incorporating dynamic cosine graph learning for robust label transmission and seamless label extraction.
• Implemented representation alignment with an independent constraint to improve cluster separability.
• Utilized trace ratio for self-balanced clustering, minimizing the need for hyperparameter tuning.

Main Results:

• The IB2MC method demonstrates linear time complexity concerning sample size, requiring no pre- or post-processing for joint optimization.
• IB2MC achieved superior average performance, ranking first across twelve real-world datasets against seventeen state-of-the-art baseline methods.

Conclusions:

• IB2MC offers a secure, synergistic, and simple-to-tune framework for multiview clustering.
• The proposed method effectively balances information compression and task-relevant information retention, outperforming existing approaches in unsupervised clustering tasks.