Why PageRank beat RotatE on our UC drug ranking task
A knowledge-graph embedding lost to a two-minute graph walk. Here's what that actually means.
- ml
- knowledge-graphs
- drug-repurposing
We trained RotatE on a biomedical knowledge graph to rank drugs for ulcerative colitis. Then we ran Personalized PageRank from the UC nodes — no training, full graph, laptop runtime.
PageRank won.
That isn't a dunk on embeddings for their own sake. It's a reminder that when your graph already encodes treatment-adjacent structure, "learning" can rediscover proximity. A strong null baseline matters more than a fancier scoring function.
Full writeup coming. This stub is here so the blog isn't empty while the long version cooks.