What They Don’t Tell You About Training GNNs: Lessons Learned from a Link Prediction Benchmark Study
Galadriel Brière
IBDM, Marseille
Date(s) : 18/11/2026 iCal
15h00 - 16h00
Graph Neural Networks (GNNs) are increasingly used to learn from relational data such as biomedical knowledge graphs, where they can help predict missing interactions between biological entities. In practice though, training a GNN that performs well is not that straightforward. Drawing on a benchmark study of data leakage in biomedical link prediction, this talk covers the practical lessons I’ve learned along the way. Going from how GNNs work, to architectural choices, training, evaluation and interpretation of results, I will walk through the process of training a GNN for link prediction and guide you around the pitfalls I came across, so you know where to look when your own model misbehaves.
Emplacement
I2M Luminy - TPR2, Amphithéâtre Herbrand 130-134 (1er étage)
Catégories



