Skip to main content
Search

# GraphPDHG

1 related articles

Subscribe to updates
  1. arXiv

    GraphPDHG: a PDHG-aligned message-passing network solves graph saddle-point problems, accelerating convergence and improving size generalization

    The authors introduce GraphPDHG, a node-edge message-passing graph neural network inspired by the Chambolle-Pock Primal-Dual Hybrid Gradient (PDHG) method for general graph saddle-point problems; they prove a single layer can exactly implement one preconditioned PDHG step, that linearized dynamics reduce to preconditioned graph-Laplacian dynamics without memory and recover heavy-ball acceleration with edge memory, and show on convex clustering that the model provides better warm starts for the second-order solver SSNAL and achieves lower primal objective values than GCN and GAT baselines as graph size grows.

You’re all caught up