Gradient-Free Feature Propagation for Missing-Attribute Reconstruction in Cold-Start Graphs
When only some graph nodes carry feature vectors, the missing attributes can be reconstructed by iteratively diffusing known features across the graph without any gradient training. Redefining the propagation's boundary conditions as a soft penalty (rather than hard-resetting known nodes) and adding virtual edges to connect isolated low-degree nodes turns the update into a contraction that converges quickly and mitigates the cold-start problem for low-degree nodes.
2501.00743
Attribute-missing graphs are graphs where only some nodes carry feature vectors; reconstructing the rest is needed before graph neural networks can classify nodes. Feature Propagation reconstructs th…