Wasserstein Discriminant Dictionary Learning for Graph Representation.

Area of Science:

• Graph representation learning
• Topological data analysis
• Machine learning

Background:

• Discriminant learning on graphs faces challenges in computing scatters and characterizing complex topologies.
• Existing methods struggle with the high dimensionality and variability of graph structures.

Purpose of the Study:

• To propose a novel Wasserstein Discriminant Dictionary Learning (WDDL) framework for enhanced graph representation.
• To address limitations in discriminant learning and robust graph topology modeling.
• To facilitate graph-based pattern analysis tasks like classification and retrieval.

Main Methods:

• Constructing a graph dictionary with representative graph samples (graph keys).
• Developing a Wasserstein Graph Representation (WGR) process using graph dictionary lookup.
• Introducing a Wasserstein discriminant loss (WD-loss) with optimizable graph keys.
• Implementing a joint-Wasserstein graph embedding module with Kron-Gromov-Wasserstein (KGW) metric for topology mining.

Main Results:

• The WDDL framework effectively models complex graph topologies and achieves discriminant learning.
• The proposed KGW metric robustly captures cross-graph connection patterns.
• Experimental results validate the framework's effectiveness in graph classification and cross-modal retrieval.

Conclusions:

• WDDL offers a robust and effective approach to graph representation learning and pattern analysis.
• The framework successfully overcomes challenges in discriminant learning and topological characterization.
• WDDL demonstrates significant potential for various graph-based machine learning applications.