An Epistemic Position-Based Click Model: From Interactions to Epistemic Distributions of Relevance and Bias

Research Question and Approach

Search and recommendation systems rely on click models to disentangle item relevance from position bias. Traditional models provide only frequentist point estimates, failing to quantify the model's confidence in its predictions. This paper addresses this limitation by proposing an epistemic position-based click model (PBM). Instead of predicting single values, the model uses evidential deep learning to output parameters for Beta distributions, representing the epistemic uncertainty of both relevance and position bias. Because the number of query-item pairs is vast, the authors use a neural network to predict these distribution parameters, while using a lookup table for position-specific bias.

Challenges in Optimization

Optimizing an epistemic model requires integrating over the parameter space, which is computationally expensive and prone to numerical instability. The authors identify that standard Monte Carlo gradient estimation suffers from high variance and underflow issues due to the extremely small likelihood values involved. To solve this, they introduce a numerically stable log-likelihood estimation using the log-sum-exp trick and propose a variance-reduction technique by conditioning the likelihood estimation on partial samples of position bias.

By incorporating Bayesian uncertainty into click modeling, this work provides a foundation for more robust off-policy evaluation and learning-to-rank. Standard models often treat all predictions with equal confidence, which can lead to risky decisions in downstream tasks when data is sparse or noisy. This epistemic approach allows systems to distinguish between well-supported preferences and high-uncertainty estimates, offering a more reliable way to handle long-tail items and previously unseen queries.