Reproduce “Rethinking Link Prediction for Directed Graphs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingguo He, Yuhe Guo, Yanping Zheng, Zhewei Wei, Stephan Günnemann, Xiaokui Xiao
Researchers from Renmin University of China, Technical University of Munich, and National University of Singapore introduced DirLinkBench, a standardized benchmark for directed link prediction that mitigates common evaluation flaws like label leakage. They also proposed SDGAE, a novel spectral graph auto-encoder, which achieved top performance on their new benchmark, outperforming existing methods on 4 of 7 datasets.
Key insights
- Dual embedding methods exhibit greater expressiveness for preserving asymmetry and reconstructing directed graphs compared to single-embedding approaches.
- Existing directed link prediction benchmarks are often unreliable, with issues like label leakage artificially inflating reported performance and simple MLPs often outperforming complex GNNs in flawed setups.
- Performance in directed link prediction is significantly influenced by factors such as feature inputs (e.g., in/out degrees), choice of loss function (BCE outperforms CE), decoder design, and negative sampling strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.