Transductive Sharpening boosts node classification by using unlabeled predictions

In a new preprint, Brown Zaz, Mar Gonzàlez I Català, Ferran Hernandez Caralt, Moshe Eliasof, and Pietro Liò revisit the training objective for semi-supervised node classification on graphs. Their key insight: in the transductive setting (where the full graph is observed but only some nodes have labels), models produce predictions for every node during training. Standard supervised losses discard predictions on unlabeled nodes because no ground truth exists. The authors instead use prediction confidence—specifically, they minimize prediction entropy on unlabeled nodes while counterbalancing with a modified loss on labeled nodes. This simple loss-level modification, called Transductive Sharpening (TS), requires no changes to the backbone architecture.

TS is inspired by the decomposition of cross-entropy into a label-dependent alignment term and a label-independent entropy term. By reducing entropy on unlabeled nodes, the model becomes sharper and more confident, effectively leveraging the latent structure of unlabeled predictions. The team evaluated TS across a wide range of node classification benchmarks and observed consistent performance improvements. The method is architecture-agnostic, easy to implement (code provided), and orthogonal to ongoing innovations in graph neural network design. It suggests that even without architectural advances, merely rethinking the loss function can yield substantial gains in semi-supervised learning on graphs.