New GNN study reveals when cross-task transfer works and fails

A new paper from Neelam Akula, Surbhi Kumar, Murat Kantarcioglu, and Baris Coskunuzer (arXiv:2607.28525, ICML 2026) tackles a messy problem in graph machine learning: when you have one graph but multiple tasks, can you reuse supervision between node classification (NC) and link prediction (LP)? The answer, based on experiments across three GNN backbones (GCN, GraphSAGE, GPS), is a cautious yes with sharp caveats. The authors introduce a leakage-free evaluation protocol that fixes node and edge splits, uses a shared message-passing graph excluding evaluated edges, and uses fixed negatives for LP. This protocol aims to eliminate the incompatible splits and negative sampling tricks that made prior conclusions unreliable. Under this controlled setup, they find that transfer is strongly directional: NC→LP consistently boosts performance on homophilic graphs, where connected nodes share labels. That direction works like a supervised warm start for link prediction.

The reverse direction, LP→NC, is far more fragile. Naively reusing link prediction representations can actually degrade node classification accuracy. It only becomes reliably positive in a “structure-dominant regime” where LP is easy but NC is unsaturated — effectively making LP act as structural pretraining. To help practitioners decide when to share an encoder, the authors propose the CoTask Score (CTS), which summarizes joint NC+LP utility across both tasks. They also show that simple dataset statistics, especially graph homophily, can predict which transfer mechanism will work, helping teams avoid negative transfer in production. The paper is 17 pages and includes code/data links via the arXiv listing. For engineers building multi-task GNN pipelines, this is a useful reminder that representation reuse is not symmetric — and that choosing the wrong transfer direction can silently hurt accuracy.