DAISY LabSharif University of Technology
Publications

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 preprint, 14 April 2025 Preprint

arXiv

Abstract

Most well-known recommendation models employ representation learning to map users and items into a unified embedding space for matching assessment. These approaches have primary limitations, especially with explicit feedback and sparse data: proneness to overfitting and failure to incorporate epistemic uncertainty in predictions. We propose a Bayesian Deep Ensemble Collaborative Filtering method named BDECF. To improve model generalisation and quality, we use Bayesian neural networks, which incorporate uncertainty within their weight parameters, and introduce an interpretable non-linear matching approach for user and item embeddings leveraging the attention mechanism. We further endorse an ensemble-based supermodel to generate more robust and reliable predictions. Extensive experiments and ablation studies across public real-world datasets with differing sparsity confirm the method's effectiveness.

In short

Recommenders trained on sparse explicit feedback are confidently wrong in exactly the places where they have seen the least data. BDECF addresses that directly: weights are Bayesian, so the model carries epistemic uncertainty, and an ensemble of such models produces predictions that can be trusted or discounted according to how much the members disagree.

The matching function between user and item embeddings is attention-based and non-linear, which also makes the match interpretable.

Cite this paper

@misc{cheraghi2025epistemic,
  title = {{Epistemic Uncertainty-aware Recommendation Systems via Bayesian Deep Ensemble Learning}},
  author = {Radin Cheraghi and Amir Mohammad Mahfoozi and Sepehr Zolfaghari and Mohammadshayan Shabani and Maryam Ramezani and Hamid R. Rabiee},
  year = {2025},
  eprint = {2504.10753},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.10753}
}

Related work