This thesis tackles a core challenge in real-world classification and ranking tasks where only positives are known and the rest are unlabelled: how to construct reliable negatives without introducing bias. This work focuses on proposing a Positive–Unlabeled (PU) learning framework that uses genetic algorithms (GAs) to automatically discover effective strategies for reliable-negative from the unlabeled pool and for configuring the downstream model.
The thesis delivers: (1) developing a PU learning model using genetic algorithms; (2) a reproducible open-source implementation with configuration-as-code; (3) a rigorous benchmark against common PU baselines under shared splits and statistical testing.
Together, these results show how evolutionary search can turn PU's labeling constraints into a principled, tunable pipeline.
If you want to learn more or get involved, please contact Pooya Zakeri .