Granular Information Bottleneck for Deep Multi-Modal Clustering.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Deep multi-modal clustering aims to improve accuracy by integrating information from diverse data sources.
• Current methods often overlook the importance of relationships across different data granularity levels, impacting performance.
• Integrating information across granularities is crucial for robust feature representation in multi-modal clustering.

Purpose of the Study:

• To propose a novel granular information bottleneck (GIB) for deep multi-modal clustering.
• To address the limitations of existing methods by incorporating multi-granularity information integration.
• To enhance feature representation learning for improved clustering accuracy and reliability.

Main Methods:

• Introduced a dual-tiered information bottleneck mechanism operating at both granular and sample levels.
• Employed adaptive representation of sample points using granular balls at varying granularity levels.
• Utilized information compression and preservation to exploit modality complementarity and optimize cluster alignment, with variational optimization for convergence.

Main Results:

• The proposed GIB model effectively captures feature distributions within clusters by considering multi-granularity information.
• GIB demonstrates enhanced inter-cluster separability through discriminative feature representations.
• Experimental results validate the superior accuracy and reliability of the GIB model compared to existing methods.

Conclusions:

• The granular information bottleneck (GIB) offers a significant advancement in deep multi-modal clustering.
• Considering multi-granularity relationships is essential for improving clustering performance.
• GIB provides a reliable and accurate approach for multi-modal data analysis.