Graph Convolutional Neural Networks With Diverse Negative Samples via Decomposed Determinant Point Processes.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Artificial Intelligence

Background:

• Graph convolutional neural networks (GCNs) excel at graph representation learning via message passing.
• Current GCNs primarily use adjacent nodes as positive samples, neglecting valuable information from non-adjacent nodes.
• Over-smoothing is a common issue in GCNs, limiting representation quality.

Purpose of the Study:

• To develop a novel approach for graph representation learning by incorporating diverse negative samples.
• To improve the efficiency and effectiveness of GCNs by addressing limitations in sample selection.
• To mitigate the over-smoothing problem in graph neural networks.

Main Methods:

• Utilized quality-diversity decomposition in determinantal point processes (DPPs) to generate diverse negative samples from non-neighboring nodes.
• Incorporated both graph structure and node representations into the negative sample distribution.
• Proposed a new shortest-path-based method to enhance the computational efficiency of DPP sampling.

Main Results:

• The proposed method significantly improved the performance of standard representation learning on node classification tasks.
• The integration of diverse negative samples effectively alleviated the over-smoothing problem in GCNs.
• Empirical evaluations demonstrated the superiority of the new approach over existing methods.

Conclusions:

• The novel method of incorporating diverse negative samples enhances GCN performance and tackles over-smoothing.
• The shortest-path-based DPP sampling offers a computationally efficient alternative.
• This research provides a promising direction for advancing graph representation learning.