Uncertainty-aware recommendation with Bayesian deep ensembles
BDECF combines Bayesian neural networks with deep ensembles so a recommender can report how certain it is, not only what it predicts.
Recommenders trained on sparse explicit feedback are most confident exactly where they have the least evidence. BDECF carries epistemic uncertainty in its weights and aggregates an ensemble, so disagreement between members becomes a usable signal.
People: Maryam Ramezani
Related publication
Epistemic Uncertainty-aware Recommendation Systems via Bayesian Deep Ensemble Learning
Radin Cheraghi, Amir Mohammad Mahfoozi, Sepehr Zolfaghari, Mohammadshayan Shabani, Maryam Ramezani, Hamid R. Rabiee
arXiv|2025|Preprint