Multi-View Clustering via Triplex Information Maximization.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multi-view clustering (MVC) aims to leverage information from multiple data sources for improved clustering.
• Existing MVC methods often struggle to effectively integrate diverse information across views.

Purpose of the Study:

• To propose Triplex Information Maximization (TIM), a novel framework for multi-view clustering.
• To enhance clustering by integrating view-specific, cross-view feature-level, and cross-view cluster-level information.

Main Methods:

• TIM is based on three principles: contained, complementary, and compatible information maximization.
• An automatic view correlation learning (AVCL) mechanism quantifies complementary information across views.
• Two TIM versions, TIM-F (feature-based) and TIM-C (cluster-based), are proposed, optimized via a two-stage method.

Main Results:

• TIM effectively utilizes self-contained, complementary, and compatible information within and among views.
• The AVCL mechanism automatically learns cross-view weights, outperforming methods with fixed or view-specific weights.
• Extensive experiments demonstrate the superiority of TIM-F and TIM-C over state-of-the-art MVC methods.

Conclusions:

• TIM provides a robust framework for multi-view clustering by maximizing mutual information across different levels.
• The proposed AVCL mechanism offers an effective approach for adaptive integration of complementary information.
• TIM-F and TIM-C represent significant advancements in multi-view clustering performance.