SIGIR 2026 paper introduces first evidential click model with epistemic uncertainty

User interactions with search rankings are influenced by both item relevance and display position—a longstanding challenge in information retrieval. Traditional click models treat these factors as fixed point estimates, offering no indication of prediction confidence. This paper introduces the first evidential deep-learning alternative, building on the position-based click model. The model processes item and position features through a neural network that outputs beta-distributions for each relevance and position-bias variable. These distributions encode epistemic uncertainty—the model's uncertainty about its own predictions—enabling practitioners to gauge how much trust to place in each estimate. The authors tackle optimization challenges specific to evidential learning, proposing approximation and conditioning techniques to ensure numerical stability and variance reduction.

Experimental results demonstrate that the proposed model successfully captures epistemic uncertainty on previously unseen data, while standard policy gradient methods fail to learn meaningful distributions. This is a significant departure from frequentist approaches; by outputting entire distributions instead of single values, the model provides richer information for downstream tasks like ranking optimization and A/B testing. The work, published at SIGIR 2026, represents a principled integration of Bayesian uncertainty into click modeling. For practitioners, this means more reliable inferences about user behavior and better-informed decisions when adapting search or recommendation systems. The code and data are available alongside the paper.