arXiv subjects

Richard Mudd

Publications and source records attributed to Richard Mudd.

7 recordsLinked to original sources

arXiv2026

Breaking the Winner's Curse with Bayesian Hybrid Shrinkage

The widespread adoption of randomized controlled trials (A/B Tests) for decision-making has introduced a pervasive "Winner's Curse": experiments selected for launch often exhibit upwardly biased effect estimates and invalid confidence intervals. This selection bias leads to over-optimistic impact projections and undermines decision-making, particularly in low-power regimes. We propose Bayesian Hybrid Shrinkage (BHS), an empirical Bayes (EB) framework that leverages data-driven priors to mitigate selection bias and provides accurate uncertainty quantification. Unlike traditional EB methods that apply uniform shrinkage, BHS introduces an experiment-specific "local" shrinkage factor that incorporates individual experiment characteristics, improving robustness against prior misspecification. We also derive a closed-form inference strategy designed for high-throughput production environments. Extensive simulations and real-world evaluations at Meta Platforms demonstrate that BHS outperforms existing methods in terms of bias reduction and interval coverage, even under substantial violations of modeling assumptions.

stat.ME↗

arXiv2025

Breaking the Winner's Curse with Bayesian Hybrid Shrinkage

A 'Winner's Curse' arises in large-scale online experimentation platforms when the same experiments are used to both select treatments and evaluate their effects. In these settings, classical difference-in-means estimators of treatment effects are upwardly biased and conventional confidence intervals are rendered invalid. The bias scales with the magnitude of sampling variability and the selection threshold, and inversely with the treatment's true effect size. We propose a new Bayesian approach that incorporates experiment-specific 'local shrinkage' factors that mitigate sensitivity to the choice of prior and improve robustness to assumption violations. We demonstrate how the associated posterior distribution can be estimated without numerical integration techniques, making it a practical choice for at-scale deployment. Through simulation, we evaluate the performance of our approach under various scenarios and find that it performs well even when assumptions about the sampling and selection processes are violated. In an empirical evaluation, our approach demonstrated superior performance over alternative methods, providing more accurate estimates with well-calibrated uncertainty quantification.

stat.ME↗

arXiv2025

Bayesian Predictive Probabilities for Online Experimentation

The widespread adoption of online randomized controlled experiments (A/B Tests) for decision-making has created ongoing capacity constraints which necessitate interim analyses. As a consequence, platform users are increasingly motivated to use ad-hoc means of optimizing limited resources via peeking. Such processes, however, are error prone and often misaligned with end-of-experiment outcomes (e.g., inflated type-I error). We introduce a system based on Bayesian Predictive Probabilities that enable us to perform interim analyses without compromising fidelity of the experiment; This idea has been widely utilized in applications outside of the technology domain to more efficiently make decisions in experiments. Motivated by at-scale deployment within an experimentation platform, we demonstrate how predictive probabilities can be estimated without numerical integration techniques and recommend systems to study its properties at scale as an ongoing health check, along with system design recommendations - all on experiment data from Instagram - to demonstrate practical benefits that it enables.

stat.AP↗

arXiv2025

Treatment Effect Learning Under Sequential Randomization

Sequential treatment assignments in online experiments lead to complex dependency structures, often rendering identification, estimation and inference over treatments a challenge. Treatments in one session (e.g., a user logging on) can have an effect that persists into subsequent sessions, leading to cumulative effects on outcomes measured at a later stage. This can render standard methods for identification and inference trivially misspecified. We propose T-Learners layered into the G-Formula for this setting, building on literature from causal machine learning and identification in sequential settings. In a simple simulation, this approach prevents decaying accuracy in the presence of carry-over effects, highlighting the importance of identification and inference strategies tailored to the nature of systems often seen in the tech domain.

stat.AP↗

arXiv2023

Active learning with biased non-response to label requests

Active learning can improve the efficiency of training prediction models by identifying the most informative new labels to acquire. However, non-response to label requests can impact active learning's effectiveness in real-world contexts. We conceptualise this degradation by considering the type of non-response present in the data, demonstrating that biased non-response is particularly detrimental to model performance. We argue that biased non-response is likely in contexts where the labelling process, by nature, relies on user interactions. To mitigate the impact of biased non-response, we propose a cost-based correction to the sampling strategy--the Upper Confidence Bound of the Expected Utility (UCB-EU)--that can, plausibly, be applied to any active learning algorithm. Through experiments, we demonstrate that our method successfully reduces the harm from labelling non-response in many settings. However, we also characterise settings where the non-response bias in the annotations remains detrimental under UCB-EU for specific sampling methods and data generating processes. Finally, we evaluate our method on a real-world dataset from an e-commerce platform. We show that UCB-EU yields substantial performance improvements to conversion models that are trained on clicked impressions. Most generally, this research serves to both better conceptualise the interplay between types of non-response and model improvements via active learning, and to provide a practical, easy-to-implement correction that mitigates model degradation.

cs.LG↗

arXiv2023

TCE: A Test-Based Approach to Measuring Calibration Error

This paper proposes a new metric to measure the calibration error of probabilistic binary classifiers, called test-based calibration error (TCE). TCE incorporates a novel loss function based on a statistical test to examine the extent to which model predictions differ from probabilities estimated from data. It offers (i) a clear interpretation, (ii) a consistent scale that is unaffected by class imbalance, and (iii) an enhanced visual representation with repect to the standard reliability diagram. In addition, we introduce an optimality criterion for the binning procedure of calibration error metrics based on a minimal estimation error of the empirical probabilities. We provide a novel computational algorithm for optimal bins under bin-size constraints. We demonstrate properties of TCE through a range of experiments, including multiple real-world imbalanced datasets and ImageNet 1000.

stat.ML↗

arXiv2023

Explaining Predictive Uncertainty with Information Theoretic Shapley Values

Researchers in explainable artificial intelligence have developed numerous methods for helping users understand the predictions of complex supervised learning models. By contrast, explaining the $\textit{uncertainty}$ of model outputs has received relatively little attention. We adapt the popular Shapley value framework to explain various types of predictive uncertainty, quantifying each feature's contribution to the conditional entropy of individual model outputs. We consider games with modified characteristic functions and find deep connections between the resulting Shapley values and fundamental quantities from information theory and conditional independence testing. We outline inference procedures for finite sample error rate control with provable guarantees, and implement efficient algorithms that perform well in a range of experiments on real and simulated data. Our method has applications to covariate shift detection, active learning, feature selection, and active feature-value acquisition.

stat.ML↗