Multi-View Learning a Decomposable Affinity Matrix via Tensor Self-Representation on Grassmann Manifold.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view clustering leverages cross-view information to partition data into categories.
• Existing methods often struggle to effectively integrate information from multiple views, leading to suboptimal clustering.
• Previous tensor-based approaches focused on latent tensor representation, treating affinity matrices as byproducts.

Purpose of the Study:

• To propose a novel multi-view clustering method that directly learns a well-structured affinity matrix.
• To address limitations of existing methods by avoiding the destruction of intrinsic clustering structures.
• To enhance clustering performance by jointly optimizing for an integrative subspace and a consensus affinity matrix.

Main Methods:

• Employed a tensor learning model to unify multiple feature spaces into a latent low-rank tensor space.
• Utilized the Grassmann manifold to merge individual views, obtaining an integrative subspace and a consensus affinity matrix.
• Modeled the process with a unified objective function, jointly optimizing for decomposable affinity matrix mining.

Main Results:

• The proposed method demonstrated superior performance compared to popular multi-view clustering techniques.
• Experiments conducted on eight real-world datasets validated the effectiveness of the approach.
• The joint optimization strategy successfully mined a decomposable affinity matrix, preserving clustering structures.

Conclusions:

• The novel method effectively addresses the challenge of learning a well-structured affinity matrix in multi-view clustering.
• Integrating tensor learning with Grassmann manifold provides a robust framework for multi-view data analysis.
• The approach offers significant improvements in clustering accuracy by preserving intrinsic data relationships.