Graph Neural Network Meets Sparse Representation: Graph Sparse Neural Networks via Exclusive Group Lasso

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph Neural Networks (GNNs) commonly aggregate all neighborhood information, making them vulnerable to structural noise like redundant connections.
• This noise can degrade the performance and robustness of GNN models.

Purpose of the Study:

• To introduce a novel GNN approach that mitigates the impact of structural noise in graph data.
• To enhance the reliability and performance of GNNs through sparse aggregation techniques.

Main Methods:

• Proposed Graph Sparse Neural Networks (GSNNs) leveraging Sparse Representation (SR) theory for selective neighbor aggregation.
• Developed a continuous relaxation model, Exclusive Group Lasso GNNs (EGLassoGNNs), to address the optimization challenges of discrete sparse constraints.
• Derived an effective algorithm for optimizing the EGLassoGNNs model.

Main Results:

• The proposed EGLassoGNNs model demonstrated superior performance compared to existing methods.
• Experimental results confirmed the enhanced robustness of EGLassoGNNs against structural noise in graph datasets.
• The model effectively selects reliable neighbors for message aggregation, improving GNN accuracy.

Conclusions:

• EGLassoGNNs offer a more robust and performant alternative to traditional GNNs, particularly in noisy graph environments.
• The integration of sparse representation theory provides a principled way to handle structural noise.
• The developed optimization algorithm effectively enables the practical application of these sparse GNNs.