Less for Better: A View Filter-Driven Graph Representation Fusion Network.

Area of Science:

• Graph Representation Learning
• Multi-view Learning
• Machine Learning

Background:

• Multi-view learning enhances graph representation by fusing complementary information.
• Existing methods often fail to address noise introduced by irrelevant views, degrading performance.
• Irrelevant views can negatively impact the quality of graph representations.

Purpose of the Study:

• To propose a novel multi-view representation learning framework, ViFi, that filters informative views and discards irrelevant ones.
• To enhance graph representation quality by addressing the issue of noisy or irrelevant views.
• To improve the performance of graph-based tasks like classification and clustering.

Main Methods:

• Developed ViFi, a View Filter-driven graph representation fusion network.
• Designed an entropy-based adaptive view filter to dynamically select informative views based on feature-topology entropy.
• Implemented an optimized fusion mechanism using a novel information gain function to integrate filtered views.

Main Results:

• ViFi effectively filters irrelevant views, reducing noise and enhancing view complementarity.
• The proposed framework demonstrates superior performance in graph classification and clustering tasks.
• ViFi significantly outperforms existing state-of-the-art multi-view graph representation learning approaches.

Conclusions:

• ViFi offers an effective solution for handling irrelevant views in multi-view graph representation learning.
• The framework's view filtering and optimized fusion mechanisms lead to improved representation quality.
• ViFi provides a robust approach for enhancing performance in graph-based machine learning applications.