Rethinking Link Prediction for Directed Graphs.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Directed link prediction is vital for applications like recommendation systems and network analysis.
• Existing Graph Neural Network (GNN) methods lack expressiveness analysis and standardized benchmarks.
• Current benchmarks hinder fair evaluation of directed link prediction techniques.

Purpose of the Study:

• To propose a unified framework for assessing the expressiveness of directed link prediction methods.
• To introduce DirLinkBench, a novel benchmark for standardized and robust evaluation.
• To develop an improved model, Spectral Directed Graph Auto-Encoder (SDGAE), for enhanced performance.

Main Methods:

• Developed a unified framework to analyze dual embeddings and decoder designs in GNNs for link prediction.
• Introduced DirLinkBench, a comprehensive benchmark with standardized evaluation protocols.
• Proposed and theoretically revisited the Spectral Directed Graph Auto-Encoder (SDGAE) model.

Main Results:

• Current GNN methods show limitations in directed link prediction performance on DirLinkBench.
• DiGAE demonstrated superior performance compared to other baselines.
• SDGAE achieved state-of-the-art average performance on the DirLinkBench benchmark.

Conclusions:

• The proposed framework and DirLinkBench offer valuable tools for evaluating directed link prediction methods.
• SDGAE represents a significant advancement in directed link prediction, outperforming existing approaches.
• Further research is needed to address key factors and open challenges in the field.