Rumi Chunara

Miao Zhang, Harvineet Singh, Lazarus Chok et al.

The increasing availability of high-resolution satellite imagery has enabled the use of machine learning to support land-cover measurement and inform policy-making. However, labelling satellite images is expensive and is available for only some locations. This prompts the use of transfer learning to adapt models from data-rich locations to others. Given the potential for high-impact applications of satellite imagery across geographies, a systematic assessment of transfer learning implications is warranted. In this work, we consider the task of land-cover segmentation and study the fairness implications of transferring models across locations. We leverage a large satellite image segmentation benchmark with 5987 images from 18 districts (9 urban and 9 rural). Via fairness metrics we quantify disparities in model performance along two axes -- across urban-rural locations and across land-cover classes. Findings show that state-of-the-art models have better overall accuracy in rural areas compared to urban areas, through unsupervised domain adaptation methods transfer learning better to urban versus rural areas and enlarge fairness gaps. In analysis of reasons for these findings, we show that raw satellite images are overall more dissimilar between source and target districts for rural than for urban locations. This work highlights the need to conduct fairness analysis for satellite imagery segmentation models and motivates the development of methods for fair transfer learning in order not to introduce disparities between places, particularly urban and rural locations.

Dongkyu Cho, Miao Zhang, Rumi Chunara

Clinical foundation models are evaluated with factual or exam-style medical QA, but treatment decisions must change when patient context changes. We introduce ClinPivot, an auditable treatment-decision benchmark built from biomedical relations and pivoted patient contexts. ClinPivot asks whether models change treatment choices when new clinical constraints shift the action space. We find that strong medical QA performance does not reliably predict decision-making performance: frontier models and task-adapted Qwen variants often fail to change decisions correctly, and model rankings shift across evaluation regimes. Decision-structured supervision improves pivot-sensitive decision-making and medical QA under matched knowledge budgets, while lightweight replay reduces losses in general assistant ability.

Martin Marek, Dongkyu Cho, Shikai Qiu et al.

Models trained on a new task typically degrade on prior tasks, a phenomenon known as forgetting. Traditionally, mitigating forgetting has required replaying stored exemplars from prior tasks, which is often impractical. By contrast, language models can sample from their own training distribution, and we show that these self-generated samples serve as effective replay data, nearly eliminating forgetting. We find that forgetting nonetheless persists when the model has little remaining capacity: models pretrained close to saturation cannot absorb new information without overwriting prior knowledge. When capacity is not the limiting factor, low learning rates reduce forgetting but require substantially more training steps. Replay breaks this tradeoff, enabling fast, high-learning-rate finetuning without forgetting.

Miao Zhang, Rumi Chunara

Satellite imagery is being leveraged for many societally critical tasks across climate, economics, and public health. Yet, because of heterogeneity in landscapes (e.g. how a road looks in different places), models can show disparate performance across geographic areas. Given the important potential of disparities in algorithmic systems used in societal contexts, here we consider the risk of urban-rural disparities in identification of land-cover features. This is via semantic segmentation (a common computer vision task in which image regions are labelled according to what is being shown) which uses pre-trained image representations generated via contrastive self-supervised learning. We propose fair dense representation with contrastive learning (FairDCL) as a method for de-biasing the multi-level latent space of convolution neural network models. The method improves feature identification by removing spurious model representations which are disparately distributed across urban and rural areas, and is achieved in an unsupervised way by contrastive pre-training. The obtained image representation mitigates downstream urban-rural prediction disparities and outperforms state-of-the-art baselines on real-world satellite images. Embedding space evaluation and ablation studies further demonstrate FairDCL's robustness. As generalizability and robustness in geographic imagery is a nascent topic, our work motivates researchers to consider metrics beyond average accuracy in such applications.

Dongkyu Cho, Xiyue Li, Samrachana Adhikari et al.

Continual learning aims to update models under distribution shift without forgetting, yet many high-stakes deployments, such as healthcare, also require interpretability. In practice, models that adapt well (e.g., deep networks) are often opaque, while models that are interpretable (e.g., decision trees) are brittle under shift, making it difficult to achieve both properties simultaneously. In response, we propose Tree of Concepts, an interpretable continual learning framework that uses a shallow decision tree to define a fixed, rule-based concept interface and trains a concept bottleneck model to predict these concepts from raw features. Continual updates act on the concept extractor and label head while keeping concept semantics stable over time, yielding explanations that do not drift across sequential updates. On multiple tabular healthcare benchmarks under continual learning protocols, our method achieves a stronger stability-plasticity trade-off than existing baselines, including replay-enhanced variants. Our results suggest that structured concept interfaces can support continual adaptation while preserving a consistent audit interface in non-stationary, high-stakes domains.

Miao Zhang, Kelly Chen, Md Mehrab Tanjim et al.

Large Language Model (LLM) outputs often vary across user sociodemographic attributes, leading to disparities in factual accuracy, utility, and safety, even for objective questions where demographic information is irrelevant. Unlike prior work on stereotypical or representational bias, this paper studies identity-dependent degradation of core response quality. We show empirically that such degradation arises from biased generation behavior, despite factual knowledge being robustly encoded across identities. Motivated by this mismatch, we propose a lightweight, training-free framework for identity-robust generation that selectively neutralizes non-critical identity information while preserving semantically essential attributes, thus maintaining output content integrity. Experiments across four benchmarks and 18 sociodemographic identities demonstrate an average 77% reduction in identity-dependent bias compared to vanilla prompting and a 45% reduction relative to prompt-based defenses. Our work addresses a critical gap in mitigating the impact of user identity cues in prompts on core generation quality.

Salman Rahman, Lavender Yao Jiang, Saadia Gabriel et al.

Advances in large language models (LLMs) provide new opportunities in healthcare for improved patient care, clinical decision-making, and enhancement of physician and administrator workflows. However, the potential of these models importantly depends on their ability to generalize effectively across clinical environments and populations, a challenge often underestimated in early development. To better understand reasons for these challenges and inform mitigation approaches, we evaluated ClinicLLM, an LLM trained on [HOSPITAL]'s clinical notes, analyzing its performance on 30-day all-cause readmission prediction focusing on variability across hospitals and patient characteristics. We found poorer generalization particularly in hospitals with fewer samples, among patients with government and unspecified insurance, the elderly, and those with high comorbidities. To understand reasons for lack of generalization, we investigated sample sizes for fine-tuning, note content (number of words per note), patient characteristics (comorbidity level, age, insurance type, borough), and health system aspects (hospital, all-cause 30-day readmission, and mortality rates). We used descriptive statistics and supervised classification to identify features. We found that, along with sample size, patient age, number of comorbidities, and the number of words in notes are all important factors related to generalization. Finally, we compared local fine-tuning (hospital specific), instance-based augmented fine-tuning and cluster-based fine-tuning for improving generalization. Among these, local fine-tuning proved most effective, increasing AUC by 0.25% to 11.74% (most helpful in settings with limited data). Overall, this study provides new insights for enhancing the deployment of large language models in the societally important domain of healthcare, and improving their performance for broader populations.

Miao Zhang, Rumi Chunara

Performance disparities of image recognition across different demographic populations are known to exist in deep learning-based models, but previous work has largely addressed such fairness problems assuming knowledge of sensitive attribute labels. To overcome this reliance, previous strategies have involved separate learning structures to expose and adjust for disparities. In this work, we explore a new paradigm that does not require sensitive attribute labels, and evades the need for extra training by leveraging general-purpose vision-language model (VLM), as a rich knowledge source for common sensitive attributes. We analyze the correspondence between VLM predicted and human defined sensitive attribute distribution. We find that VLMs can recognize samples with clear attribute information encoded in image representations, thus capture under-performed samples conflicting with attribute-related bias. We train downstream target classifiers by re-sampling and augmenting under-performed attribute groups. Extensive experiments on multiple benchmark facial attribute classification datasets show fairness gains of the model over existing unsupervised baselines that tackle with arbitrary bias. The work indicates that vision-language models can extract discriminative sensitive information prompted by language, and be used to promote model fairness.

Miao Zhang, Salman Rahman, Vishwali Mhasawade et al.

New data sources, and artificial intelligence (AI) methods to extract information from them are becoming plentiful, and relevant to decision making in many societal applications. An important example is street view imagery, available in over 100 countries, and considered for applications such as assessing built environment aspects in relation to community health outcomes. Relevant to such uses, important examples of bias in the use of AI are evident when decision-making based on data fails to account for the robustness of the data, or predictions are based on spurious correlations. To study this risk, we utilize 2.02 million GSV images along with health, demographic, and socioeconomic data from New York City. Initially, we demonstrate that built environment characteristics inferred from GSV labels at the intra-city level may exhibit inadequate alignment with the ground truth. We also find that the average individual-level behavior of physical inactivity significantly mediates the impact of built environment features by census tract, as measured through GSV. Finally, using a causal framework which accounts for these mediators of environmental impacts on health, we find that altering 10% of samples in the two lowest tertiles would result in a 4.17 (95% CI 3.84 to 4.55) or 17.2 (95% CI 14.4 to 21.3) times bigger decrease on the prevalence of obesity or diabetes, than the same proportional intervention on the number of crosswalks by census tract. This work illustrates important issues of robustness and model specification for informing effective allocation of interventions using new data sources.

Harvineet Singh, Fan Xia, Mi-Ok Kim et al.

In fairness audits, a standard objective is to detect whether a given algorithm performs substantially differently between subgroups. Properly powering the statistical analysis of such audits is crucial for obtaining informative fairness assessments, as it ensures a high probability of detecting unfairness when it exists. However, limited guidance is available on the amount of data necessary for a fairness audit, lacking directly applicable results concerning commonly used fairness metrics. Additionally, the consideration of unequal subgroup sample sizes is also missing. In this tutorial, we address these issues by providing guidance on how to determine the required subgroup sample sizes to maximize the statistical power of hypothesis tests for detecting unfairness. Our findings are applicable to audits of binary classification models and multiple fairness metrics derived as summaries of the confusion matrix. Furthermore, we discuss other aspects of audit study designs that can increase the reliability of audit results.

Dongkyu Cho, Amy B. Z. Zhang, Bilel Fehri et al.

Chain-of-thought (CoT) reasoning exposes the intermediate thinking process of large language models (LLMs), yet verifying those traces at scale remains unsolved. In response, we introduce the idea of decision pivots-minimal, verifiable checkpoints that any correct reasoning path must visit. We hypothesize that correct reasoning, though stylistically diverse, converge on the same pivot set, while incorrect ones violate at least one pivot. Leveraging this property, we propose a self-training pipeline that (i) samples diverse reasoning paths and mines shared decision pivots, (ii) compresses each trace into pivot-focused short-path reasoning using an auxiliary verifier, and (iii) post-trains the model using its self-generated outputs. The proposed method aligns reasoning without ground truth reasoning data or external metrics. Experiments on standard benchmarks such as LogiQA, MedQA, and MATH500 show the effectiveness of our method.

Dongkyu Cho, Miao Zhang, Rumi Chunara

Data augmentation is a widely used strategy to improve model robustness and generalization by enriching training datasets with synthetic examples. While large language models (LLMs) have demonstrated strong generative capabilities for this purpose, their applications in high-stakes domains like healthcare present unique challenges due to the risk of generating clinically incorrect or misleading information. In this work, we propose a novel query-based model collaboration framework that integrates expert-level domain knowledge to guide the augmentation process to preserve critical medical information. Experiments on clinical prediction tasks demonstrate that our lightweight collaboration-based approach consistently outperforms existing LLM augmentation methods while improving safety through reduced factual errors. This framework addresses the gap between LLM augmentation potential and the safety requirements of specialized domains.

Dongkyu Cho, Rumi Chunara

Data augmentation is a promising tool for enhancing out-of-distribution generalization, where the key is to produce diverse, challenging variations of the source domain via costly targeted augmentations that maximize its generalization effect. Conversely, random augmentation is inexpensive but is deemed suboptimal due to its limited effect. In this paper, we revisit random augmentation and explore methods to address its shortcomings. We show that the stochastic nature of random augmentation can produce a set of colliding augmentations that distorts the learned features, similar to catastrophic forgetting. We propose a simple solution that improves the generalization effect of random augmentation by addressing forgetting, which displays strong generalization performance across various single source domain generalization (sDG) benchmarks.

Dongkyu Cho, Taesup Moon, Rumi Chunara et al.

Continual learning (CL) has traditionally focused on minimizing exemplar memory, a constraint often misaligned with modern systems where GPU time, not storage, is the primary bottleneck. This paper challenges this paradigm by investigating a more realistic regime: one where memory is abundant enough to mitigate forgetting, but full retraining from scratch remains prohibitively expensive. In this practical "middle ground", we find that the core challenge shifts from stability to plasticity, as models become biased toward prior tasks and struggle to learn new ones. Conversely, improved stability allows simple replay baselines to outperform the state-of-the-art methods at a fraction of the GPU cost. To address this newly surfaced trade-off, we propose Weight Space Consolidation, a lightweight method that combines (1) rank-based parameter resets to restore plasticity with (2) weight averaging to enhance stability. Validated on both class-incremental learning with image classifiers and continual instruction tuning with large language models, our approach outperforms strong baselines while matching the low computational cost of replay, offering a scalable alternative to expensive full-retraining. These findings challenge long-standing CL assumptions and establish a new, cost-efficient baseline for real-world CL systems where exemplar memory is no longer the limiting factor.

Vishwali Mhasawade, Rumi Chunara

Transported mediation effects provide an avenue to understand how upstream interventions (such as improved neighborhood conditions like green spaces) would work differently when applied to different populations as a result of factors that mediate the effects. However, when mediators are missing in the population where the effect is to be transported, these estimates could be biased. We study this issue of missing mediators, motivated by challenges in public health, wherein mediators can be missing, not at random. We propose a sensitivity analysis framework that quantifies the impact of missing mediator data on transported mediation effects. This framework enables us to identify the settings under which the conditional transported mediation effect is rendered insignificant for the subgroup with missing mediator data. Specifically, we provide the bounds on the transported mediation effect as a function of missingness. We then apply the framework to longitudinal data from the Moving to Opportunity Study, a large-scale housing voucher experiment, to quantify the effect of missing mediators on transport effect estimates of voucher receipt, an upstream intervention on living location, in childhood on subsequent risk of mental health or substance use disorder mediated through parental health across sites. Our findings provide a tangible understanding of how much missing data can be withstood for unbiased effect estimates.

Vishwali Mhasawade, Salman Rahman, Zoe Haskell-Craig et al.

Previous work has highlighted that existing post-hoc explanation methods exhibit disparities in explanation fidelity (across 'race' and 'gender' as sensitive attributes), and while a large body of work focuses on mitigating these issues at the explanation metric level, the role of the data generating process and black box model in relation to explanation disparities remains largely unexplored. Accordingly, through both simulations as well as experiments on a real-world dataset, we specifically assess challenges to explanation disparities that originate from properties of the data: limited sample size, covariate shift, concept shift, omitted variable bias, and challenges based on model properties: inclusion of the sensitive attribute and appropriate functional form. Through controlled simulation analyses, our study demonstrates that increased covariate shift, concept shift, and omission of covariates increase explanation disparities, with the effect pronounced higher for neural network models that are better able to capture the underlying functional form in comparison to linear models. We also observe consistent findings regarding the effect of concept shift and omitted variable bias on explanation disparities in the Adult income dataset. Overall, results indicate that disparities in model explanations can also depend on data and model properties. Based on this systematic investigation, we provide recommendations for the design of explanation methods that mitigate undesirable disparities.

Umang Bhatt, Javier Antorán, Yunfeng Zhang et al.

Algorithmic transparency entails exposing system properties to various stakeholders for purposes that include understanding, improving, and contesting predictions. Until now, most research into algorithmic transparency has predominantly focused on explainability. Explainability attempts to provide reasons for a machine learning model's behavior to stakeholders. However, understanding a model's specific behavior alone might not be enough for stakeholders to gauge whether the model is wrong or lacks sufficient knowledge to solve the task at hand. In this paper, we argue for considering a complementary form of transparency by estimating and communicating the uncertainty associated with model predictions. First, we discuss methods for assessing uncertainty. Then, we characterize how uncertainty can be used to mitigate model unfairness, augment decision-making, and build trustworthy systems. Finally, we outline methods for displaying uncertainty to stakeholders and recommend how to collect information required for incorporating uncertainty into existing ML pipelines. This work constitutes an interdisciplinary review drawn from literature spanning machine learning, visualization/HCI, design, decision-making, and fairness. We aim to encourage researchers and practitioners to measure, communicate, and use uncertainty as a form of transparency.

Vishwali Mhasawade, Rumi Chunara

Algorithmic systems are known to impact marginalized groups severely, and more so, if all sources of bias are not considered. While work in algorithmic fairness to-date has primarily focused on addressing discrimination due to individually linked attributes, social science research elucidates how some properties we link to individuals can be conceptualized as having causes at macro (e.g. structural) levels, and it may be important to be fair to attributes at multiple levels. For example, instead of simply considering race as a causal, protected attribute of an individual, the cause may be distilled as perceived racial discrimination an individual experiences, which in turn can be affected by neighborhood-level factors. This multi-level conceptualization is relevant to questions of fairness, as it may not only be important to take into account if the individual belonged to another demographic group, but also if the individual received advantaged treatment at the macro-level. In this paper, we formalize the problem of multi-level fairness using tools from causal inference in a manner that allows one to assess and account for effects of sensitive attributes at multiple levels. We show importance of the problem by illustrating residual unfairness if macro-level sensitive attributes are not accounted for, or included without accounting for their multi-level nature. Further, in the context of a real-world task of predicting income based on macro and individual-level attributes, we demonstrate an approach for mitigating unfairness, a result of multi-level sensitive attributes.

Vishwali Mhasawade, Yuan Zhao, Rumi Chunara

Research in population and public health focuses on the mechanisms between different cultural, social, and environmental factors and their effect on the health, of not just individuals, but communities as a whole. We present here a very brief introduction into research in these fields, as well as connections to existing machine learning work to help activate the machine learning community on such topics and highlight specific opportunities where machine learning, public and population health may synergize to better achieve health equity.

Harvineet Singh, Rina Singh, Vishwali Mhasawade et al.

We study the problem of learning fair prediction models for unseen test sets distributed differently from the train set. Stability against changes in data distribution is an important mandate for responsible deployment of models. The domain adaptation literature addresses this concern, albeit with the notion of stability limited to that of prediction accuracy. We identify sufficient conditions under which stable models, both in terms of prediction accuracy and fairness, can be learned. Using the causal graph describing the data and the anticipated shifts, we specify an approach based on feature selection that exploits conditional independencies in the data to estimate accuracy and fairness metrics for the test set. We show that for specific fairness definitions, the resulting model satisfies a form of worst-case optimality. In context of a healthcare task, we illustrate the advantages of the approach in making more equitable decisions.

Vishwali Mhasawade, Nabeel Abdur Rehman, Rumi Chunara

While machine learning is rapidly being developed and deployed in health settings such as influenza prediction, there are critical challenges in using data from one environment in another due to variability in features; even within disease labels there can be differences (e.g. "fever" may mean something different reported in a doctor's office versus in an online app). Moreover, models are often built on passive, observational data which contain different distributions of population subgroups (e.g. men or women). Thus, there are two forms of instability between environments in this observational transport problem. We first harness knowledge from health to conceptualize the underlying causal structure of this problem in a health outcome prediction task. Based on sources of stability in the model, we posit that for human-sourced data and health prediction tasks we can combine environment and population information in a novel population-aware hierarchical Bayesian domain adaptation framework that harnesses multiple invariant components through population attributes when needed. We study the conditions under which invariant learning fails, leading to reliance on the environment-specific attributes. Experimental results for an influenza prediction task on four datasets gathered from different contexts show the model can improve prediction in the case of largely unlabelled target data from a new environment and different constituent population, by harnessing both environment and population invariant information. This work represents a novel, principled way to address a critical challenge by blending domain (health) knowledge and algorithmic innovation. The proposed approach will have a significant impact in many social settings wherein who and where the data comes from matters.

Mohammad Akbari, Rumi Chunara

Blood glucose value prediction is an important task in diabetes management. While it is reported that glucose concentration is sensitive to social context such as mood, physical activity, stress, diet, alongside the influence of diabetes pathologies, we need more research on data and methodologies to incorporate and evaluate signals about such temporal context into prediction models. Person-generated data sources, such as actively contributed surveys as well as passively mined data from social media offer opportunity to capture such context, however the self-reported nature and sparsity of such data mean that such data are noisier and less specific than physiological measures such as blood glucose values themselves. Therefore, here we propose a Gaussian Process model to both address these data challenges and combine blood glucose and latent feature representations of contextual data for a novel multi-signal blood glucose prediction task. We find this approach outperforms common methods for multi-variate data, as well as using the blood glucose values in isolation. Given a robust evaluation across two blood glucose datasets with different forms of contextual information, we conclude that multi-signal Gaussian Processes can improve blood glucose prediction by using contextual information and may provide a significant shift in blood glucose prediction research and practice.

Nabeel Abdur Rehman, Umar Saif, Rumi Chunara

The global population at risk of mosquito-borne diseases such as dengue, yellow fever, chikungunya and Zika is expanding. Infectious disease models commonly incorporate environmental measures like temperature and precipitation. Given increasing availability of high-resolution satellite imagery, here we consider including landscape features from satellite imagery into infectious disease prediction models. To do so, we implement a Convolutional Neural Network (CNN) model trained on Imagenet data and labelled landscape features in satellite data from London. We then incorporate landscape features from satellite image data from Pakistan, labelled using the CNN, in a well-known Susceptible-Infectious-Recovered epidemic model, alongside dengue case data from 2012-2016 in Pakistan. We study improvement of the prediction model for each of the individual landscape features, and assess the feasibility of using image labels from a different place. We find that incorporating satellite-derived landscape features can improve prediction of outbreaks, which is important for proactive and strategic surveillance and control programmes.

Mohammad Akbari, Kunal Relia, Anas Elghafari et al.

Online communities provide a unique way for individuals to access information from those in similar circumstances, which can be critical for health conditions that require daily and personalized management. As these groups and topics often arise organically, identifying the types of topics discussed is necessary to understand their needs. As well, these communities and people in them can be quite diverse, and existing community detection methods have not been extended towards evaluating these heterogeneities. This has been limited as community detection methodologies have not focused on community detection based on semantic relations between textual features of the user-generated content. Thus here we develop an approach, NeuroCom, that optimally finds dense groups of users as communities in a latent space inferred by neural representation of published contents of users. By embedding of words and messages, we show that NeuroCom demonstrates improved clustering and identifies more nuanced discussion topics in contrast to other common unsupervised learning approaches.

Vishwali Mhasawade, Nabeel Abdur Rehman, Rumi Chunara

Population attributes are essential in health for understanding who the data represents and precision medicine efforts. Even within disease infection labels, patients can exhibit significant variability; "fever" may mean something different when reported in a doctor's office versus from an online app, precluding directly learning across different datasets for the same prediction task. This problem falls into the domain adaptation paradigm. However, research in this area has to-date not considered who generates the data; symptoms reported by a woman versus a man, for example, could also have different implications. We propose a novel population-aware domain adaptation approach by formulating the domain adaptation task as a multi-source hierarchical Bayesian framework. The model improves prediction in the case of largely unlabelled target data by harnessing both domain and population invariant information.

Nabeel Abdur Rehman, Maxwell Matthaios Aliapoulios, Disha Umarwani et al.

Acute respiratory infections have epidemic and pandemic potential and thus are being studied worldwide, albeit in many different contexts and study formats. Predicting infection from symptom data is critical, though using symptom data from varied studies in aggregate is challenging because the data is collected in different ways. Accordingly, different symptom profiles could be more predictive in certain studies, or even symptoms of the same name could have different meanings in different contexts. We assess state-of-the-art transfer learning methods for improving prediction of infection from symptom data in multiple types of health care data ranging from clinical, to home-visit as well as crowdsourced studies. We show interesting characteristics regarding six different study types and their feature domains. Further, we demonstrate that it is possible to use data collected from one study to predict infection in another, at close to or better than using a single dataset for prediction on itself. We also investigate in which conditions specific transfer learning and domain adaptation methods may perform better on symptom data. This work has the potential for broad applicability as we show how it is possible to transfer learning from one public health study design to another, and data collected from one study may be used for prediction of labels for another, even collected through different study designs, populations and contexts.