Temporal Positive-Unlabeled Learning for Biomedical Hypothesis Generation via Risk Estimation

Abstract

We treat biomedical hypothesis generation as a positive-unlabeled learning problem: known valid connections are positive examples, but unobserved connections could be either true negatives or undiscovered positives. Temporal risk estimation over the evolving graph allows us to rank candidate hypotheses by their probability of becoming confirmed.

Published at NeurIPS 2020 — the highest-profile venue in my publication record.