| Accurate probability calibration is crucial in click-through rate (CTR) prediction, because miscalibrated models can lead to suboptimal bidding strategies and degraded user experience. Existing calibration approaches are almost exclusively applied post-hoc, and as a result, these methods fail to integrate into an end-to-end learning framework to facilitate improved online calibration. We introduce FABR, a Field-aware Bayesian Conjugacy-based Regularization, a novel approach that directly constrains models' predictions to better align with distributions consistent with the user's empirical click behavior at a field-level, by applying lightweight Beta-Bernoulli priors to the original loss function. We illustrate the effectiveness of our approach using a large eBay dataset and the public Avazu dataset, across six CTR models and three ranking losses. By adding even a small level of FABR, we demonstrate an average relative improvement of 4.86% in field-level ECE and 9.40% in field-level MCE on the eBay dataset, with an average relative AUC change of less than 0.2%. These results show that embedding Bayesian conjugacy as a structural constraint is an effective and model-agnostic way to achieve improved field-level calibration, making FABR directly applicable to large-scale online advertising and other probability-critical domains. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.