Emma Pierson

Charvi Rastogi, Ivan Stelmakh, Alina Beygelzimer et al. · cmu, microsoft-research

How do author perceptions match up to the outcomes of the peer-review process and perceptions of others? In a top-tier computer science conference (NeurIPS 2021) with more than 23,000 submitting authors and 9,000 submitted papers, we survey the authors on three questions: (i) their predicted probability of acceptance for each of their papers, (ii) their perceived ranking of their own papers based on scientific contribution, and (iii) the change in their perception about their own papers after seeing the reviews. The salient results are: (1) Authors have roughly a three-fold overestimate of the acceptance probability of their papers: The median prediction is 70% for an approximately 25% acceptance rate. (2) Female authors exhibit a marginally higher (statistically significant) miscalibration than male authors; predictions of authors invited to serve as meta-reviewers or reviewers are similarly calibrated, but better than authors who were not invited to review. (3) Authors' relative ranking of scientific contribution of two submissions they made generally agree (93%) with their predicted acceptance probabilities, but there is a notable 7% responses where authors think their better paper will face a worse outcome. (4) The author-provided rankings disagreed with the peer-review decisions about a third of the time; when co-authors ranked their jointly authored papers, co-authors disagreed at a similar rate -- about a third of the time. (5) At least 30% of respondents of both accepted and rejected papers said that their perception of their own paper improved after the review process. The stakeholders in peer review should take these findings into account in setting their expectations from peer review.

J. D. Zamfirescu-Pereira, Jerry Chen, Emily Wen et al. · berkeley

Algorithms provide powerful tools for detecting and dissecting human bias and error. Here, we develop machine learning methods to to analyze how humans err in a particular high-stakes task: image interpretation. We leverage a unique dataset of 16,135,392 human predictions of whether a neighborhood voted for Donald Trump or Joe Biden in the 2020 US election, based on a Google Street View image. We show that by training a machine learning estimator of the Bayes optimal decision for each image, we can provide an actionable decomposition of human error into bias, variance, and noise terms, and further identify specific features (like pickup trucks) which lead humans astray. Our methods can be applied to ensure that human-in-the-loop decision-making is accurate and fair and are also applicable to black-box algorithmic systems.

Rajiv Movva, Divya Shanmugam, Kaihua Hou et al.

Healthcare data in the United States often records only a patient's coarse race group: for example, both Indian and Chinese patients are typically coded as "Asian." It is unknown, however, whether this coarse coding conceals meaningful disparities in the performance of clinical risk scores across granular race groups. Here we show that it does. Using data from 418K emergency department visits, we assess clinical risk score performance disparities across 26 granular groups for three outcomes, five risk scores, and four performance metrics. Across outcomes and metrics, we show that the risk scores exhibit significant granular performance disparities within coarse race groups. In fact, variation in performance within coarse groups often *exceeds* the variation between coarse groups. We explore why these disparities arise, finding that outcome rates, feature distributions, and the relationships between features and outcomes all vary significantly across granular groups. Our results suggest that healthcare providers, hospital systems, and machine learning researchers should strive to collect, release, and use granular race data in place of coarse race data, and that existing analyses may significantly underestimate racial disparities in performance.

Serina Chang, Alicja Chaszczewicz, Emma Wang et al.

Generating social networks is essential for many applications, such as epidemic modeling and social simulations. The emergence of generative AI, especially large language models (LLMs), offers new possibilities for social network generation: LLMs can generate networks without additional training or need to define network parameters, and users can flexibly define individuals in the network using natural language. However, this potential raises two critical questions: 1) are the social networks generated by LLMs realistic, and 2) what are risks of bias, given the importance of demographics in forming social ties? To answer these questions, we develop three prompting methods for network generation and compare the generated networks to a suite of real social networks. We find that more realistic networks are generated with "local" methods, where the LLM constructs relations for one persona at a time, compared to "global" methods that construct the entire network at once. We also find that the generated networks match real networks on many characteristics, including density, clustering, connectivity, and degree distribution. However, we find that LLMs emphasize political homophily over all other types of homophily and significantly overestimate political homophily compared to real social networks.

Rajiv Movva, Sidhika Balachandar, Kenny Peng et al.

Large language models (LLMs) are dramatically influencing AI research, spurring discussions on what has changed so far and how to shape the field's future. To clarify such questions, we analyze a new dataset of 16,979 LLM-related arXiv papers, focusing on recent trends in 2023 vs. 2018-2022. First, we study disciplinary shifts: LLM research increasingly considers societal impacts, evidenced by 20x growth in LLM submissions to the Computers and Society sub-arXiv. An influx of new authors -- half of all first authors in 2023 -- are entering from non-NLP fields of CS, driving disciplinary expansion. Second, we study industry and academic publishing trends. Surprisingly, industry accounts for a smaller publication share in 2023, largely due to reduced output from Google and other Big Tech companies; universities in Asia are publishing more. Third, we study institutional collaboration: while industry-academic collaborations are common, they tend to focus on the same topics that industry focuses on rather than bridging differences. The most prolific institutions are all US- or China-based, but there is very little cross-country collaboration. We discuss implications around (1) how to support the influx of new authors, (2) how industry trends may affect academics, and (3) possible effects of (the lack of) collaboration.

Pang Wei Koh, Shiori Sagawa, Henrik Marklund et al.

Distribution shifts -- where the training distribution differs from the test distribution -- can substantially degrade the accuracy of machine learning (ML) systems deployed in the wild. Despite their ubiquity in the real-world deployments, these distribution shifts are under-represented in the datasets widely used in the ML community today. To address this gap, we present WILDS, a curated benchmark of 10 datasets reflecting a diverse range of distribution shifts that naturally arise in real-world applications, such as shifts across hospitals for tumor identification; across camera traps for wildlife monitoring; and across time and location in satellite imaging and poverty mapping. On each dataset, we show that standard training yields substantially lower out-of-distribution than in-distribution performance. This gap remains even with models trained by existing methods for tackling distribution shifts, underscoring the need for new methods for training models that are more robust to the types of distribution shifts that arise in practice. To facilitate method development, we provide an open-source package that automates dataset loading, contains default model architectures and hyperparameters, and standardizes evaluations. Code and leaderboards are available at https://wilds.stanford.edu.

Shuvom Sadhuka, Sophia Lin, Emma Pierson et al.

Many sequential decision settings in healthcare feature funnel structures characterized by a series of stages, such as screenings or evaluations, where the number of patients who advance to each stage progressively decreases and decisions become increasingly costly. For example, an oncologist may first conduct a breast exam, followed by a mammogram for patients with concerning exams, followed by a biopsy for patients with concerning mammograms. A key challenge is that the ground truth outcome, such as the biopsy result, is only revealed at the end of this funnel. The selective censoring of the ground truth can introduce statistical biases in risk estimation, especially in underserved patient groups, whose outcomes are more frequently censored. We develop a Bayesian model for funnel decision structures, drawing from prior work on selective labels and censoring. We first show in synthetic settings that our model is able to recover the true parameters and predict outcomes for censored patients more accurately than baselines. We then apply our model to a dataset of emergency department visits, where in-hospital mortality is observed only for those who are admitted to either the hospital or ICU. We find that there are gender-based differences in hospital and ICU admissions. In particular, our model estimates that the mortality risk threshold to admit women to the ICU is higher for women (5.1%) than for men (4.5%).

Rajiv Movva, Kenny Peng, Nikhil Garg et al.

We describe HypotheSAEs, a general method to hypothesize interpretable relationships between text data (e.g., headlines) and a target variable (e.g., clicks). HypotheSAEs has three steps: (1) train a sparse autoencoder on text embeddings to produce interpretable features describing the data distribution, (2) select features that predict the target variable, and (3) generate a natural language interpretation of each feature (e.g., "mentions being surprised or shocked") using an LLM. Each interpretation serves as a hypothesis about what predicts the target variable. Compared to baselines, our method better identifies reference hypotheses on synthetic datasets (at least +0.06 in F1) and produces more predictive hypotheses on real datasets (~twice as many significant findings), despite requiring 1-2 orders of magnitude less compute than recent LLM-based methods. HypotheSAEs also produces novel discoveries on two well-studied tasks: explaining partisan differences in Congressional speeches and identifying drivers of engagement with online headlines.

Kenny Peng, Rajiv Movva, Jon Kleinberg et al.

While sparse autoencoders (SAEs) have generated significant excitement, a series of negative results have added to skepticism about their usefulness. Here, we establish a conceptual distinction that reconciles competing narratives surrounding SAEs. We argue that while SAEs may be less effective for acting on known concepts, SAEs are powerful tools for discovering unknown concepts. This distinction cleanly separates existing negative and positive results, and suggests several classes of SAE applications. Specifically, we outline use cases for SAEs in (i) ML interpretability, explainability, fairness, auditing, and safety, and (ii) social and health sciences.

Sidhika Balachandar, Nikhil Garg, Emma Pierson

Machine learning models are often trained to predict the outcome resulting from a human decision. For example, if a doctor decides to test a patient for disease, will the patient test positive? A challenge is that historical decision-making determines whether the outcome is observed: we only observe test outcomes for patients doctors historically tested. Untested patients, for whom outcomes are unobserved, may differ from tested patients along observed and unobserved dimensions. We propose a Bayesian model class which captures this setting. The purpose of the model is to accurately estimate risk for both tested and untested patients. Estimating this model is challenging due to the wide range of possibilities for untested patients. To address this, we propose two domain constraints which are plausible in health settings: a prevalence constraint, where the overall disease prevalence is known, and an expertise constraint, where the human decision-maker deviates from purely risk-based decision-making only along a constrained feature set. We show theoretically and on synthetic data that domain constraints improve parameter inference. We apply our model to a case study of cancer risk prediction, showing that the model's inferred risk predicts cancer diagnoses, its inferred testing policy captures known public health policies, and it can identify suboptimalities in test allocation. Though our case study is in healthcare, our analysis reveals a general class of domain constraints which can improve model estimation in many settings.

Gabriel Agostini, Emma Pierson, Nikhil Garg

Decision-makers often observe the occurrence of events through a reporting process. City governments, for example, rely on resident reports to find and then resolve urban infrastructural problems such as fallen street trees, flooded basements, or rat infestations. Without additional assumptions, there is no way to distinguish events that occur but are not reported from events that truly did not occur--a fundamental problem in settings with positive-unlabeled data. Because disparities in reporting rates correlate with resident demographics, addressing incidents only on the basis of reports leads to systematic neglect in neighborhoods that are less likely to report events. We show how to overcome this challenge by leveraging the fact that events are spatially correlated. Our framework uses a Bayesian spatial latent variable model to infer event occurrence probabilities and applies it to storm-induced flooding reports in New York City, further pooling results across multiple storms. We show that a model accounting for under-reporting and spatial correlation predicts future reports more accurately than other models, and further induces a more equitable set of inspections: its allocations better reflect the population and provide equitable service to non-white, less traditionally educated, and lower-income residents. This finding reflects heterogeneous reporting behavior learned by the model: reporting rates are higher in Census tracts with higher populations, proportions of white residents, and proportions of owner-occupied households. Our work lays the groundwork for more equitable proactive government services, even with disparate reporting behavior.

Erica Chiang, Divya Shanmugam, Ashley N. Beecy et al.

Disease progression models are widely used to inform the diagnosis and treatment of many progressive diseases. However, a significant limitation of existing models is that they do not account for health disparities that can bias the observed data. To address this, we develop an interpretable Bayesian disease progression model that captures three key health disparities: certain patient populations may (1) start receiving care only when their disease is more severe, (2) experience faster disease progression even while receiving care, or (3) receive follow-up care less frequently conditional on disease severity. We show theoretically and empirically that failing to account for any of these disparities can result in biased estimates of severity (e.g., underestimating severity for disadvantaged groups). On a dataset of heart failure patients, we show that our model can identify groups that face each type of health disparity, and that accounting for these disparities while inferring disease severity meaningfully shifts which patients are considered high-risk.

Hyewon Jeong, Sarah Jabbour, Yuzhe Yang et al. · uw

The third ML4H symposium was held in person on December 10, 2023, in New Orleans, Louisiana, USA. The symposium included research roundtable sessions to foster discussions between participants and senior researchers on timely and relevant topics for the \ac{ML4H} community. Encouraged by the successful virtual roundtables in the previous year, we organized eleven in-person roundtables and four virtual roundtables at ML4H 2022. The organization of the research roundtables at the conference involved 17 Senior Chairs and 19 Junior Chairs across 11 tables. Each roundtable session included invited senior chairs (with substantial experience in the field), junior chairs (responsible for facilitating the discussion), and attendees from diverse backgrounds with interest in the session's topic. Herein we detail the organization process and compile takeaways from these roundtable discussions, including recent advances, applications, and open challenges for each topic. We conclude with a summary and lessons learned across all roundtables. This document serves as a comprehensive review paper, summarizing the recent advancements in machine learning for healthcare as contributed by foremost researchers in the field.

Divya Shanmugam, Shuvom Sadhuka, Manish Raghavan et al.

It remains difficult to evaluate machine learning classifiers in the absence of a large, labeled dataset. While labeled data can be prohibitively expensive or impossible to obtain, unlabeled data is plentiful. Here, we introduce Semi-Supervised Model Evaluation (SSME), a method that uses both labeled and unlabeled data to evaluate machine learning classifiers. SSME is the first evaluation method to take advantage of the fact that: (i) there are frequently multiple classifiers for the same task, (ii) continuous classifier scores are often available for all classes, and (iii) unlabeled data is often far more plentiful than labeled data. The key idea is to use a semi-supervised mixture model to estimate the joint distribution of ground truth labels and classifier predictions. We can then use this model to estimate any metric that is a function of classifier scores and ground truth labels (e.g., accuracy or expected calibration error). We present experiments in four domains where obtaining large labeled datasets is often impractical: (1) healthcare, (2) content moderation, (3) molecular property prediction, and (4) image annotation. Our results demonstrate that SSME estimates performance more accurately than do competing methods, reducing error by 5.1x relative to using labeled data alone and 2.4x relative to the next best competing method. SSME also improves accuracy when evaluating performance across subsets of the test distribution (e.g., specific demographic subgroups) and when evaluating the performance of language models.

Sidhika Balachandar, Shuvom Sadhuka, Bonnie Berger et al.

Graph neural networks (GNNs) are widely used in urban spatiotemporal forecasting, such as predicting infrastructure problems. In this setting, government officials wish to know in which neighborhoods incidents like potholes or rodent issues occur. The true state of incidents (e.g., street conditions) for each neighborhood is observed via government inspection ratings. However, these ratings are only conducted for a sparse set of neighborhoods and incident types. We also observe the state of incidents via crowdsourced reports, which are more densely observed but may be biased due to heterogeneous reporting behavior. First, for such settings, we propose a multiview, multioutput GNN-based model that uses both unbiased rating data and biased reporting data to predict the true latent state of incidents. Second, we investigate a case study of New York City urban incidents and collect, standardize, and make publicly available a dataset of 9,615,863 crowdsourced reports and 1,041,415 government inspection ratings over 3 years and across 139 types of incidents. Finally, we show on both real and semi-synthetic data that our model can better predict the latent state compared to models that use only reporting data or models that use only rating data, especially when rating data is sparse and reports are predictive of ratings. We also quantify demographic biases in crowdsourced reporting, e.g., higher-income neighborhoods report problems at higher rates. Our analysis showcases a widely applicable approach for latent state prediction using heterogeneous, sparse, and biased data.

Emily Alsentzer, Marie-Laure Charpignon, Bill Chen et al.

The 6th Annual Conference on Health, Inference, and Learning (CHIL 2025), hosted by the Association for Health Learning and Inference (AHLI), was held in person on June 25-27, 2025, at the University of California, Berkeley, in Berkeley, California, USA. As part of this year's program, we hosted Research Roundtables to catalyze collaborative, small-group dialogue around critical, timely topics at the intersection of machine learning and healthcare. Each roundtable was moderated by a team of senior and junior chairs who fostered open exchange, intellectual curiosity, and inclusive engagement. The sessions emphasized rigorous discussion of key challenges, exploration of emerging opportunities, and collective ideation toward actionable directions in the field. In total, eight roundtables were held by 19 roundtable chairs on topics of "Explainability, Interpretability, and Transparency," "Uncertainty, Bias, and Fairness," "Causality," "Domain Adaptation," "Foundation Models," "Learning from Small Medical Data," "Multimodal Methods," and "Scalable, Translational Healthcare Solutions."

Rajiv Movva, Pang Wei Koh, Emma Pierson

Do LLMs align with human perceptions of safety? We study this question via annotation alignment, the extent to which LLMs and humans agree when annotating the safety of user-chatbot conversations. We leverage the recent DICES dataset (Aroyo et al., 2023), in which 350 conversations are each rated for safety by 112 annotators spanning 10 race-gender groups. GPT-4 achieves a Pearson correlation of $r = 0.59$ with the average annotator rating, \textit{higher} than the median annotator's correlation with the average ($r=0.51$). We show that larger datasets are needed to resolve whether LLMs exhibit disparities in how well they correlate with different demographic groups. Also, there is substantial idiosyncratic variation in correlation within groups, suggesting that race & gender do not fully capture differences in alignment. Finally, we find that GPT-4 cannot predict when one demographic group finds a conversation more unsafe than another.

Smitha Milli, Emma Pierson, Nikhil Garg

Many recommender systems optimize a linear weighting of different user behaviors, such as clicks, likes, and shares. We analyze the optimal choice of weights from the perspectives of both users and content producers who strategically respond to the weights. We consider three aspects of each potential behavior: value-faithfulness (how well a behavior indicates whether the user values the content), strategy-robustness (how hard it is for producers to manipulate the behavior), and noisiness (how much estimation error there is in predicting the behavior). Our theoretical results show that for users, up-weighting more value-faithful and less noisy behaviors leads to higher utility, while for producers, up-weighting more value-faithful and strategy-robust behaviors leads to higher welfare (and the impact of noise is non-monotonic). Finally, we apply our framework to design weights on Facebook, using a large-scale dataset of approximately 70 million URLs shared on Facebook. Strikingly, we find that our user-optimal weight vector (a) delivers higher user value than a vector not accounting for variance; (b) also enhances broader societal outcomes, reducing misinformation and raising the quality of the URL domains, outcomes that were not directly targeted in our theoretical framework.

Pang Wei Koh, Thao Nguyen, Yew Siang Tang et al.

We seek to learn models that we can interact with using high-level concepts: if the model did not think there was a bone spur in the x-ray, would it still predict severe arthritis? State-of-the-art models today do not typically support the manipulation of concepts like "the existence of bone spurs", as they are trained end-to-end to go directly from raw input (e.g., pixels) to output (e.g., arthritis severity). We revisit the classic idea of first predicting concepts that are provided at training time, and then using these concepts to predict the label. By construction, we can intervene on these concept bottleneck models by editing their predicted concept values and propagating these changes to the final prediction. On x-ray grading and bird identification, concept bottleneck models achieve competitive accuracy with standard end-to-end models, while enabling interpretation in terms of high-level clinical concepts ("bone spurs") or bird attributes ("wing color"). These models also allow for richer human-model interaction: accuracy improves significantly if we can correct model mistakes on concepts at test time.

Bo Liu, Shuyang Shi, Yongshang Wu et al.

Predicting pregnancy has been a fundamental problem in women's health for more than 50 years. Previous datasets have been collected via carefully curated medical studies, but the recent growth of women's health tracking mobile apps offers potential for reaching a much broader population. However, the feasibility of predicting pregnancy from mobile health tracking data is unclear. Here we develop four models -- a logistic regression model, and 3 LSTM models -- to predict a woman's probability of becoming pregnant using data from a women's health tracking app, Clue by BioWink GmbH. Evaluating our models on a dataset of 79 million logs from 65,276 women with ground truth pregnancy test data, we show that our predicted pregnancy probabilities meaningfully stratify women: women in the top 10% of predicted probabilities have a 89% chance of becoming pregnant over 6 menstrual cycles, as compared to a 27% chance for women in the bottom 10%. We develop a technique for extracting interpretable time trends from our deep learning models, and show these trends are consistent with previous fertility research. Our findings illustrate the potential that women's health tracking data offers for predicting pregnancy on a broader population; we conclude by discussing the steps needed to fulfill this potential.

Emma Pierson, Pang Wei Koh, Tatsunori Hashimoto et al.

Modeling how individuals evolve over time is a fundamental problem in the natural and social sciences. However, existing datasets are often cross-sectional with each individual observed only once, making it impossible to apply traditional time-series methods. Motivated by the study of human aging, we present an interpretable latent-variable model that learns temporal dynamics from cross-sectional data. Our model represents each individual's features over time as a nonlinear function of a low-dimensional, linearly-evolving latent state. We prove that when this nonlinear function is constrained to be order-isomorphic, the model family is identifiable solely from cross-sectional data provided the distribution of time-independent variation is known. On the UK Biobank human health dataset, our model reconstructs the observed data while learning interpretable rates of aging associated with diseases, mortality, and aging risk factors.

Emma Pierson, Sam Corbett-Davies, Sharad Goel

Threshold tests have recently been proposed as a useful method for detecting bias in lending, hiring, and policing decisions. For example, in the case of credit extensions, these tests aim to estimate the bar for granting loans to white and minority applicants, with a higher inferred threshold for minorities indicative of discrimination. This technique, however, requires fitting a complex Bayesian latent variable model for which inference is often computationally challenging. Here we develop a method for fitting threshold tests that is two orders of magnitude faster than the existing approach, reducing computation from hours to minutes. To achieve these performance gains, we introduce and analyze a flexible family of probability distributions on the interval [0, 1] -- which we call discriminant distributions -- that is computationally efficient to work with. We demonstrate our technique by analyzing 2.7 million police stops of pedestrians in New York City.