Jerome P. Reiter

Olivier Binette, Youngsoo Baek, Siddharth Engineer et al.

Entity resolution (record linkage, microclustering) systems are notoriously difficult to evaluate. Looking for a needle in a haystack, traditional evaluation methods use sophisticated, application-specific sampling schemes to find matching pairs of records among an immense number of non-matches. We propose an alternative that facilitates the creation of representative, reusable benchmark data sets without necessitating complex sampling schemes. These benchmark data sets can then be used for model training and a variety of evaluation tasks. Specifically, we propose an entity-centric data labeling methodology that integrates with a unified framework for monitoring summary statistics, estimating key performance metrics such as cluster and pairwise precision and recall, and analyzing root causes for errors. We validate the framework in an application to inventor name disambiguation and through simulation studies. Software: https://github.com/OlivierBinette/er-evaluation/

Olivier Binette, Jerome P. Reiter

Commonly, AI or machine learning (ML) models are evaluated on benchmark datasets. This practice supports innovative methodological research, but benchmark performance can be poorly correlated with performance in real-world applications -- a construct validity issue. To improve the validity and practical usefulness of evaluations, we propose using an estimands framework adapted from international clinical trials guidelines. This framework provides a systematic structure for inference and reporting in evaluations, emphasizing the importance of a well-defined estimation target. We illustrate our proposal on examples of commonly used evaluation methodologies - involving cross-validation, clustering evaluation, and LLM benchmarking - that can lead to incorrect rankings of competing models (rank reversals) with high probability, even when performance differences are large. We demonstrate how the estimands framework can help uncover underlying issues, their causes, and potential solutions. Ultimately, we believe this framework can improve the validity of evaluations through better-aligned inference, and help decision-makers and model users interpret reported results more effectively.

John Abowd, Lorenzo Alvisi, Cynthia Dwork et al.

Government statistical agencies collect enormously valuable data on the nation's population and business activities. Wide access to these data enables evidence-based policy making, supports new research that improves society, facilitates training for students in data science, and provides resources for the public to better understand and participate in their society. These data also affect the private sector. For example, the Employment Situation in the United States, published by the Bureau of Labor Statistics, moves markets. Nonetheless, government agencies are under increasing pressure to limit access to data because of a growing understanding of the threats to data privacy and confidentiality. "De-identification" - stripping obvious identifiers like names, addresses, and identification numbers - has been found inadequate in the face of modern computational and informational resources. Unfortunately, the problem extends even to the release of aggregate data statistics. This counter-intuitive phenomenon has come to be known as the Fundamental Law of Information Recovery. It says that overly accurate estimates of too many statistics can completely destroy privacy. One may think of this as death by a thousand cuts. Every statistic computed from a data set leaks a small amount of information about each member of the data set - a tiny cut. This is true even if the exact value of the statistic is distorted a bit in order to preserve privacy. But while each statistical release is an almost harmless little cut in terms of privacy risk for any individual, the cumulative effect can be to completely compromise the privacy of some individuals.