Searcharxiv⌕ Search

arXiv subjects

Julia Stoyanovich

Publications and source records attributed to Julia Stoyanovich.

At least 19 recordsLinked to original sources

Rules, Choice, and Chance: A Welfare Evaluation Framework for Student-School Matching

Centralized student-school matching systems, adopted in the United States, Chile, the Netherlands, and the United Kingdom, assign students to schools based on their submitted preference rankings and schools' priority policies. Yet welfare in these systems is typically measured through a narrow lens: a few aggregate statistics published by education authorities, such as the share of students unmatched or assigned to a top-ranked choice. A single aggregate cannot reveal how outcomes vary across students, or why. Welfare depends on three interacting forces: the rules policymakers set (algorithm, tie-breaking, list length limits); the choices students make (which and how many schools to rank); and chance (lottery draws and students' residential location). A rule that reduces unmatched rates may simultaneously lower top-ranked assignment rates for specific groups, and evaluating such trade-offs in advance is challenging because individual-level preference data are rarely available. We present Trace (Trade-offs in Rules, Agency, Chance, and Equity), an open-source framework that synthesizes individual-level preference profiles from aggregate data using mixtures of Mallows models and evaluates alternative rules across multiple welfare metrics, disaggregating outcomes by choice and chance. We apply Trace to two real-world systems: NYC high school matching, where we study minimum list length requirements, and Chilean school assignment, where we compare the current multiple tie-breaking rule with a counterfactual single tie-breaking rule. Both reveal substantial, heterogeneous welfare trade-offs, underscoring the value of simulation-based evaluation before implementing school choice reforms.

cs.CY↗

GRAIL: From Saliency to Semantics in Time-Series Explanation

Explaining a time-series model requires both locating the evidence and extracting meaningful knowledge from it. Yet a single timestep is not a semantic unit: the meaningful patterns live across many samples, not in any one of them. Per-timestep saliency is therefore noisy, high-dimensional, and unstable. Existing methods address this by introducing fixed-size windows and treating them as independent players in a cooperative game. Evaluating a coalition of such players requires defining a ``baseline'' region, which prior work tackles with fixed counterfactuals (e.g., a flat line, noise, or a spectral surrogate). This choice frequently dominates the resulting explanation, obscuring the model's true decisions. To address these issues, we propose GRAIL, a region-based time-series explanation framework that uses domain-specific intervals, such as the P-wave and the QRS complex of an electrocardiogram (ECG) beat, as players. We describe each region with human-readable features and generate counterfactual baselines with an on-manifold conditional generator. Such a choice of players also admits non-temporal features: static covariates enter as singleton players attributed with the temporal regions on one scale, so a patient's tabular evidence (e.g., age, sex) is weighed directly against the signal. We attribute the predictions to the features via Owen values with common-random-number variance reduction. Experimental results on both synthetic and real data show that the proposed framework produces faithful and robust explanations, evaluated under matched on- and off-manifold operators, offering meaningful knowledge to domain experts.

cs.AI↗

Explanation Multiplicity in SHAP: Characterization and Assessment

SHAP explanations are widely used in high-stakes settings to justify decisions, yet they can differ substantially across repeated runs, even when the model, the input instance, and the prediction are held fixed. Prior work has documented disagreement between explanation methods; we show that substantial disagreement arises even within SHAP across reruns of the same estimator on the same trained model and instance. We call this phenomenon explanation multiplicity and develop an evaluation methodology for characterizing it under deployment-realistic computational budgets, combining a dual-seed protocol that compares model-induced and explainer-induced variability, a hierarchy of magnitude-based, rank-based, and set-based metrics, and randomized Dirichlet and Mallows null models that provide reference scales for observed disagreement. Across multiple datasets, models, and sampling strategies, we find that explanation multiplicity is pervasive and persists even for high-confidence predictions. The relative contribution of each source depends on the data regime and model: model-induced disagreement is generally greater on smaller datasets, while explainer-induced disagreement is greater on larger datasets. Commonly used L2 distance can understate this instability, while rank-based metrics reveal substantial changes in top-ranked features, including the leading feature. Improved sampling methods such as CTE do not eliminate rank-level multiplicity, and K-Means reduces run-to-run variation while its compressed-background explanations can diverge from the empirical-distribution reference. Practitioners should treat single-run SHAP outputs as realizations of a distribution rather than as authoritative artifacts.

cs.LG↗

SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

cs.CL↗

Table2Image: Lightweight Tabular Learning with Generated Proxy Representations and Reliability Diagnostics

Deep tabular models should ideally balance predictive performance, parameter efficiency, and robustness to imperfect learning signals---properties that are rarely considered jointly. We present Table2Image, a lightweight tabular learning model built around a learned generation pathway that maps tabular inputs into intermediate, structured proxy representations. We additionally examine a variant with variance inflation factor (VIF)-informed initialization, which downweights highly collinear features at the start of training. Across datasets from OpenML-CC18 and TabZilla, Table2Image achieves competitive clean predictive performance while remaining compact relative to several large-scale neural baselines. We further introduce a unified, severity-controlled evaluation protocol under three imperfect learning conditions---irrelevant inputs, corrupted supervision, and unstable shortcut associations---that combines performance-based robustness measures with realization-level instability diagnostics for reliability characterization. Table2Image maintains a favorable balance of performance, robustness, and compactness. Controlled ablations indicate that the learned generation pathway is a key driver of the observed gains.

cs.LG↗

VirnyFlow: Optimizing ML Pipelines for Accuracy, Fairness, and Stability at Scale

Developing machine learning (ML) systems for real-world deployment requires navigating context-dependent trade-offs among accuracy, fairness, stability, and other objectives. Existing AutoML frameworks optimize pipelines efficiently, but they fix the optimization objective up front, leave it outside the developer's control during search, and rarely scale beyond a single node. We present VirnyFlow, a system that optimizes ML pipelines jointly for accuracy, fairness, and stability at scale. A user-defined evaluation protocol, with fairness measured over binary and intersectional groups, drives every layer of the optimizer: multi-objective Bayesian optimization of physical pipelines, cost-aware bandit selection of logical pipelines, and multi-criterion pruning. The architecture combines asynchronous execution over Apache Kafka with database-backed experiment management, providing fine-grained parallelism, fault tolerance, and interactive inspection of trade-offs. On six real-world datasets, VirnyFlow achieves competitive or superior performance compared to state-of-the-art AutoML systems (auto-sklearn, Alpine Meadow, FLAML) under identical resource constraints, scales to 128 workers across four nodes on datasets of up to 2.6M records, and achieves up to 7x higher speedup than the best-scaling single-node baseline, while maintaining stable accuracy and fairness as parallelism increases. A clinical case study on distribution shift and an IRB-approved user study demonstrate human-in-the-loop navigation of trade-offs in practice: rather than returning a single "best" model, VirnyFlow lets data scientists define, inspect, and iteratively refine the objectives of the search to fit their deployment context.

cs.LG↗

RelShap: Relationally Consistent Shapley Explanations

Machine learning pipelines commonly flatten relational data into single-table representations, discarding structural constraints. Widely used Shapley value-based feature attributions then rely on feature independence, evaluating the model on combinations that could never arise in the underlying data, producing misleading explanations. We propose RelShap, a framework that incorporates relational constraints and data provenance into Shapley value computation, restricting both background data and coalition evaluation to relationally valid configurations. The framework is estimator-agnostic and composes with Kernel SHAP, Monte Carlo, and Leverage SHAP without altering their sampling or weighting properties. Functional dependencies further induce equivalence classes over feature coalitions, which RelShap exploits to reduce runtime without changing Shapley values; we provide a combinatorial characterization of the expected speedup. Experiments across multiple datasets, models, and estimators show that RelShap produces explanations that are more faithful to the data-generating process, correctly identifying the dominant feature in controlled settings where existing methods, including Conditional SHAP and ManifoldShap, do not. Our code is available at: https://github.com/duneag2/relshap.

cs.LG↗

Be Fair! Can Machine Learning Engineering Agents Adhere to Fairness Constraints?

Machine learning engineering (MLE) agents promise to automate end-to-end ML pipeline development from raw data and natural language instructions, potentially making ML accessible to non-technical domain experts. However, in sensitive and regulated domains, this abstraction creates a responsibility gap: end-users may lack visibility into design choices that affect correctness, robustness, fairness, and regulatory compliance. We argue that existing benchmarks are insufficient to assess whether MLE agents can be safely applied in such settings. We propose desiderata for a responsibility-centered evaluation framework and conduct an exploratory study on melanoma classification, focusing on fairness across skin tones as a responsibility constraint. When evaluating two recent MLE agents, we find that agent-generated pipelines show high variance and consistently underperform manually designed baselines in both predictive quality and fairness, despite fairness-oriented prompts. These preliminary results suggest that further research is needed towards redesigning MLE agents to allow humans to guide the search process and reliably assess the compliance and quality of the generated ML pipelines.

cs.LG↗

"Would You Want an AI Tutor?" Understanding Stakeholder Perceptions of LLM-based Systems in the Classroom

Large Language Models (LLMs) have gained traction in educational settings, often framed as virtual tutors or teaching assistants. Following early skepticism and bans, many schools and universities have begun integrating these systems into curricula. Yet decisions about whether and how to deploy LLM-based tools are frequently made without systematic engagement with the full range of stakeholders they affect. In this paper, we argue that understanding stakeholder perceptions of LLM-based systems in the classroom is not a matter of measuring approval or acceptance, but of identifying whose concerns are surfaced, in which contexts, and with what implications for responsible design and governance. We introduce Contextualized Perceptions for the Adoption of LLMs in Education (Co-PALE), a stakeholder-first framework that connects educational context, responsible AI principles, and categories of perception to support more deliberate decision-making about the adoption of LLM-based tools. We ground Co-PALE through a targeted analysis of prior work to diagnose recurring gaps in how stakeholder perceptions are studied, and through contextually distinct educational scenarios that illustrate how the same technology raises different concerns for different stakeholders. We further examine how university faculty and K--12 parents make sense of the framework through focus groups, using their reflections to surface tensions and uncertainties. Co-PALE supports more systematic reasoning about whether, where, and for whom LLM-based tools should be deployed in education.

cs.CY↗

ExplainerPFN: Towards tabular foundation models for model-free zero-shot feature importance estimations

Computing the importance of features in supervised classification tasks is critical for model interpretability. Shapley values are a widely used approach for explaining model predictions, but require direct access to the underlying model, an assumption frequently violated in real-world deployments. We investigate whether meaningful feature attributions can be obtained in a zero-shot setting, using only the input data distribution and no evaluations of the target model. Because multiple models can produce identical predictions yet yield different Shapley decompositions, the mapping from data to attributions is not uniquely identifiable. We therefore target attributions that are "true to the data" rather than "true to the model", learning a posterior mean attribution under a meta-training prior. To this end, we introduce ExplainerPFN, a tabular foundation model built on TabPFN, pretrained on synthetic structural causal datasets supervised with exact or near-exact Shapley values, that predicts feature attributions for unseen tabular datasets without model access, gradients, or example explanations. Our contributions are fourfold: (1) we show that few-shot surrogate explainers achieve high SHAP fidelity with as few as two reference observations; (2) we propose ExplainerPFN, the first zero-shot method for estimating Shapley-value-style feature attributions without access to the underlying model or reference explanations, providing a principled attribution where no existing explainer can be applied; (3) we release an open-source implementation including the full training pipeline and synthetic data generator; and (4) through extensive experiments on real and synthetic datasets, we show that ExplainerPFN achieves performance competitive with few-shot surrogate explainers that rely on 2-10 SHAP examples.

cs.LG↗

Memory Undone: Between Knowing and Not Knowing in Data Systems

Machine learning and data systems increasingly function as infrastructures of memory: they ingest, store, and operationalize traces of personal, political, and cultural life. Yet contemporary governance demands credible forms of forgetting, from GDPR-backed deletion to harm-mitigation and the removal of manipulative content, while technical infrastructures are optimized to retain, replicate, and reuse. This work argues that "forgetting" in computational systems cannot be reduced to a single operation (e.g., record deletion) and should instead be treated as a sociotechnical practice with distinct mechanisms and consequences. We clarify a vocabulary that separates erasure (removing or disabling access to data artifacts), unlearning (interventions that bound or remove a data point influence on learned parameters and outputs), exclusion (upstream non-collection and omission), and forgetting as an umbrella term spanning agency, temporality, reversibility, and scale. Building on examples from machine unlearning, semantic dependencies in data management, participatory data modeling, and manipulation at scale, we show how forgetting can simultaneously protect rights and enable silencing. We propose reframing unlearning as a first-class capability in knowledge infrastructures, evaluated not only by compliance or utility retention, but by its governance properties: transparency, accountability, and epistemic justice.

cs.CY↗

Seasoning Data Modeling Education with GARLIC: A Participatory Co-Design Framework

Entity-Relationship (ER) modeling is commonly taught as a primarily technical activity, despite its central role in shaping how data systems represent people, processes, and institutions. Prior research in participatory design demonstrates that involving diverse stakeholders in modeling can surface tacit knowledge, challenge implicit assumptions, and produce more inclusive data representations. However, database education currently lacks structured pedagogical approaches for teaching participatory ER modeling in practice. We introduce the GARLIC methodology for teaching and learning participatory ER modeling. GARLIC adapts and extends the ONION participatory ER modeling framework of Makovska et al.(HILDA 2025) into a workshop-based learning format that combines role-playing, collaborative synthesis, guided critique, and iterative refinement. GARLIC is designed to develop both technical modeling skills and critical awareness of the social and ethical dimensions of data representation. GARLIC lowers the barrier to participatory ER modeling and equips students with practical skills for collaborative, inclusive data model design.

cs.DB↗

ShaRP: Explaining Rankings and Preferences with Shapley Values

Algorithmic decisions in critical domains such as hiring, college admissions, and lending are often based on rankings. Given the impact of these decisions on individuals, organizations, and population groups, it is essential to understand them - to help individuals improve their ranking position, design better ranking procedures, and ensure legal compliance. In this paper, we argue that explainability methods for classification and regression, such as SHAP, are insufficient for ranking tasks, and present ShaRP - Shapley Values for Rankings and Preferences - a framework that explains the contributions of features to various aspects of a ranked outcome. ShaRP computes feature contributions for various ranking-specific profit functions, such as rank and top-k, and also includes a novel Shapley value-based method for explaining pairwise preference outcomes. We provide a flexible implementation of ShaRP, capable of efficiently and comprehensively explaining ranked and pairwise outcomes over tabular data, in score-based ranking and learning-to-rank tasks. Finally, we develop a comprehensive evaluation methodology for ranking explainability methods, showing through qualitative, quantitative, and usability studies that our rank-aware QoIs offer complementary insights, scale effectively, and help users interpret ranked outcomes in practice.

cs.AI↗

Still More Shades of Null: An Evaluation Suite for Responsible Missing Value Imputation

Data missingness is a practical challenge of sustained interest to the scientific community. In this paper, we present Shades-of-Null, an evaluation suite for responsible missing value imputation. Our work is novel in two ways (i) we model realistic and socially-salient missingness scenarios that go beyond Rubin's classic Missing Completely at Random (MCAR), Missing At Random (MAR) and Missing Not At Random (MNAR) settings, to include multi-mechanism missingness (when different missingness patterns co-exist in the data) and missingness shift (when the missingness mechanism changes between training and test) (ii) we evaluate imputers holistically, based on imputation quality and imputation fairness, as well as on the predictive performance, fairness and stability of the models that are trained and tested on the data post-imputation. We use Shades-of-Null to conduct a large-scale empirical study involving 29,736 experimental pipelines, and find that while there is no single best-performing imputation approach for all missingness types, interesting trade-offs arise between predictive performance, fairness and stability, based on the combination of missingness scenario, imputer choice, and the architecture of the predictive model. We make Shades-of-Null publicly available, to enable researchers to rigorously evaluate missing value imputation methods on a wide range of metrics in plausible and socially meaningful scenarios.

cs.AI↗

An Epistemic and Aleatoric Decomposition of Arbitrariness to Constrain the Set of Good Models

Recent research reveals that machine learning (ML) models are highly sensitive to minor changes in their training procedure, such as the inclusion or exclusion of a single data point, leading to conflicting predictions on individual data points; a property termed as arbitrariness or instability in ML pipelines in prior work. Drawing from the uncertainty literature, we show that stability decomposes into epistemic and aleatoric components, capturing the consistency and confidence in prediction, respectively. We use this decomposition to provide two main contributions. Our first contribution is an extensive empirical evaluation. We find that (i) epistemic instability can be reduced with more training data whereas aleatoric instability cannot; (ii) state-of-the-art ML models have aleatoric instability as high as 79% and aleatoric instability disparities among demographic groups as high as 29% in popular fairness benchmarks; and (iii) fairness pre-processing interventions generally increase aleatoric instability more than in-processing interventions, and both epistemic and aleatoric instability are highly sensitive to data-processing interventions and model architecture. Our second contribution is a practical solution to the problem of systematic arbitrariness. We propose a model selection procedure that includes epistemic and aleatoric criteria alongside existing accuracy and fairness criteria, and show that it successfully narrows down a large set of good models (50-100 on our datasets) to a handful of stable, fair and accurate ones. We built and publicly released a python library to measure epistemic and aleatoric multiplicity in any ML pipeline alongside existing confusion-matrix-based metrics, providing practitioners with a rich suite of evaluation metrics to use to define a more precise criterion during model selection.

cs.LG↗

ONION: A Multi-Layered Framework for Participatory ER Design

We present ONION, a multi-layered framework for participatory Entity-Relationship (ER) modeling that integrates insights from design justice, participatory AI, and conceptual modeling. ONION introduces a five-stage methodology: Observe, Nurture, Integrate, Optimize, Normalize. It supports progressive abstraction from unstructured stakeholder input to structured ER diagrams. Our approach aims to reduce designer bias, promote inclusive participation, and increase transparency through the modeling process. We evaluate ONION through real-world workshops focused on sociotechnical systems in Ukraine, highlighting how diverse stakeholder engagement leads to richer data models and deeper mutual understanding. Early results demonstrate ONION's potential to host diversity in early-stage data modeling. We conclude with lessons learned, limitations and challenges involved in scaling and refining the framework for broader adoption.

cs.DB↗

We Are AI: Taking Control of Technology

Responsible AI (RAI) is the science and practice of ensuring the design, development, use, and oversight of AI are socially sustainable--benefiting diverse stakeholders while controlling the risks. Achieving this goal requires active engagement and participation from the broader public. This paper introduces "We are AI: Taking Control of Technology," a public education course that brings the topics of AI and RAI to the general audience in a peer-learning setting. We outline the goals behind the course's development, discuss the multi-year iterative process that shaped its creation, and summarize its content. We also discuss two offerings of We are AI to an active and engaged group of librarians and professional staff at New York University, highlighting successes and areas for improvement. The course materials, including a multilingual comic book series by the same name, are publicly available and can be used independently. By sharing our experience in creating and teaching We are AI, we aim to introduce these resources to the community of AI educators, researchers, and practitioners, supporting their public education efforts.

cs.CY↗

SHAP-based Explanations are Sensitive to Feature Representation

Local feature-based explanations are a key component of the XAI toolkit. These explanations compute feature importance values relative to an ``interpretable'' feature representation. In tabular data, feature values themselves are often considered interpretable. This paper examines the impact of data engineering choices on local feature-based explanations. We demonstrate that simple, common data engineering techniques, such as representing age with a histogram or encoding race in a specific way, can manipulate feature importance as determined by popular methods like SHAP. Notably, the sensitivity of explanations to feature representation can be exploited by adversaries to obscure issues like discrimination. While the intuition behind these results is straightforward, their systematic exploration has been lacking. Previous work has focused on adversarial attacks on feature-based explainers by biasing data or manipulating models. To the best of our knowledge, this is the first study demonstrating that explainers can be misled by standard, seemingly innocuous data engineering techniques.

cs.LG↗