Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors
The paper says cross-task transfer on the same graph works best from node classification to link prediction, especially on homophilic graphs.
The authors argue that earlier tests were muddied by mismatched splits, graph leakage, and inconsistent negative sampling. They propose a leakage-free NC-LP transfer protocol with fixed node and edge splits, a shared message-passing graph, and fixed negatives for link prediction. Across GCN, GraphSAGE, and GPS, LP-to-NC transfer is described as fragile unless the setting is structure-dominant. They also introduce a CoTask Score to measure joint utility when one encoder serves both tasks. ArXiv · AI/CL/LG's note
The authors argue that earlier tests were muddied by mismatched splits, graph leakage, and inconsistent negative sampling. They propose a leakage-free NC-LP transfer protocol with fixed node and edge splits, a shared message-passing graph, and fixed negatives for link prediction. Across GCN, GraphSAGE, and GPS, LP-to-NC transfer is described as fragile unless the setting is structure-dominant. They also introduce a CoTask Score to measure joint utility when one encoder serves both tasks. ArXiv · AI/CL/LG's note
score 4