MSGNN's magnetic signed Laplacian boosts learning on signed directed graphs
New spectral GNN fuses magnetic and signed Laplacians for state-of-the-art network analysis.
A research team including Yixuan He, Michael Permultter, Gesine Reinert, and Mihai Cucuringu has released MSGNN, a spectral graph neural network designed to handle two challenging network types simultaneously: signed graphs (where edges carry positive or negative sentiment) and directed graphs (where edges have direction). Their key contribution is the magnetic signed Laplacian, a matrix that naturally generalizes both the signed Laplacian (for signed graphs) and the magnetic Laplacian (for directed graphs). This allows a single GNN architecture to learn from both signed and directional information without compromising either.
In extensive experiments on node clustering and link prediction, MSGNN attains leading performance across a wide range of datasets that include purely signed, purely directed, and mixed networks. The authors also introduce a synthetic benchmark, the Signed Directed Stochastic Block Model, to enable controlled evaluation, and release several novel real-world datasets based on lead-lag relationships in financial time series. Accepted to LoG 2022, the 39-page paper offers a practical tool for analysts working with relational data that carries both polarity and direction—such as social trust networks, opinion dynamics, and financial dependency graphs.
- Introduces the magnetic signed Laplacian, unifying signed and magnetic Laplacian approaches.
- MSGNN achieves leading performance on node clustering and link prediction across signed, directed, and mixed datasets.
- Releases the Signed Directed Stochastic Block Model and new financial lead-lag datasets from real time series.
Why It Matters
Gives data scientists a practical spectral GNN for networks with both trust and direction, improving predictions in finance and social analytics.