PaperlayerView exact revision on arXiv

Source-linked AI summary

Retiring Adult: New Datasets for Fair Machine Learning

Frances Ding, Moritz Hardt, John Miller, Ludwig Schmidt

arXiv:2108.04884v3cs.LGstat.ML

TL;DR

Fairness research on tabular data relies heavily on UCI Adult, whose construction and extreme income threshold limit external validity. The paper reconstructs the dataset and releases broader Census-derived tasks spanning domains, states, and years, finding that fairness results depend strongly on context and distribution shift. These datasets provide a basis for reevaluating and extending the field’s empirical foundations.

  • Problem

    Fairness research primarily relies on UCI Adult for tabular data, despite limitations in its documentation, feature encodings, provenance, and income-threshold choice.

  • Method

    The authors reconstruct a superset of UCI Adult from Census sources and create new Census-derived prediction tasks with an accompanying folktables package.

  • Results

    The datasets reveal that fairness criteria and intervention effects vary substantially by state and income threshold, while fairness criteria are more stable over time than predictive accuracy.

  • Takeaways & Limitations

    Fairness claims and interventions should be evaluated across broader contexts, including geographic and temporal variation, rather than relying primarily on UCI Adult.

  • Takeaways & Limitations

    The work is positioned as a complement to theoretical research and seeks to reduce, rather than eliminate, overreliance on UCI Adult and benchmark data.

Abstract

from arXiv · show

Although the fairness community has recognized the importance of data, researchers in the area primarily rely on UCI Adult when it comes to tabular data. Derived from a 1994 US Census survey, this dataset has appeared in hundreds of research papers where it served as the basis for the development and comparison of many algorithmic fairness interventions. We reconstruct a superset of the UCI Adult data from available US Census sources and reveal idiosyncrasies of the UCI Adult dataset that limit its external validity. Our primary contribution is a suite of new datasets derived from US Census surveys that extend the existing data ecosystem for research on fair machine learning. We create prediction tasks relating to income, employment, health, transportation, and housing. The data span multiple years and all states of the United States, allowing researchers to study temporal shift and geographic variation. We highlight a broad initial sweep of new empirical insights relating to trade-offs between fairness criteria, performance of algorithmic interventions, and the role of distribution shift based on our new datasets. Our findings inform ongoing debates, challenge some existing narratives, and point to future research directions. Our datasets are available at https://github.com/zykls/folktables.

1 Introduction

The paper examines how reliance on UCI Adult limits the empirical foundations of fair machine learning and introduces broader Census-derived datasets. These datasets support analyses across domains, states, and years, revealing substantial contextual variation in fairness findings and intervention effects.

  • UCI Adult limitations: The authors reconstruct a superset of UCI Adult from US Census records, exposing an idiosyncrasy that limits the dataset’s external validity.The reconstruction identifies the original source and provides a basis for examining the benchmark’s origin, impact, and limitations.
  • UCI Adult limitations: UCI Adult’s $50k income threshold corresponds to the 76th US income quantile in 1994 and produces fairness findings that vary significantly with threshold choice.Violations, trade-offs between fairness criteria, and intervention effectiveness all change with the threshold; in many cases, $50k understates the broader picture.
  • New datasets: The paper releases Census-derived prediction tasks covering income, employment, health, transportation, and housing across multiple years and all US states.The accompanying folktables package provides access to these tasks and supports creating new tasks from Census data through a simple API.
  • Empirical findings: Fairness criteria and intervention effect sizes vary greatly by state, requiring statistical claims about algorithmic fairness to be qualified by context.The findings emphasize population variation as a major factor in interpreting empirical observations.
  • Empirical findings: Training in one state and testing in another generally produces unpredictable changes in accuracy and fairness, so tools may not transfer gracefully across contexts.Accuracy and fairness criteria can change in either direction under geographic distribution shift.
  • Empirical findings: Fairness criteria appear more stable over time than predictive accuracy, both before and after intervention.The paper contrasts temporal stability in fairness criteria with changes in predictive accuracy.
  • Empirical findings: Larger datasets and additional years do not necessarily reduce observed disparities, contrasting with intuition from some cognitive machine learning tasks.The authors state that increased dataset size and social progress measured in years passed do not necessarily reduce disparities.

2 Archaeology of UCI Adult: Origin, Impact, Limitations

UCI Adult became a dominant fairness benchmark, but its undocumented provenance and unusual income threshold limit the external validity of conclusions drawn from it. The authors reconstruct the dataset and test how threshold choices affect fairness interventions and model performance.

  • Impact: UCI Adult appeared in more than 300 fairness papers and became a standard test case for fairness criteria, interventions, causal modeling, and software examples.
  • Origin: The dataset’s documentation does not uniquely identify its Census source, so the authors reconstruct it from 1994 CPS ASEC data through IPUMS as IPUMS Adult.
  • Reconstruction: The reconstruction matches UCI Adult’s features using candidate CPS variables and exact row-level matches, producing 49,531 rows versus UCI Adult’s 48,842.
  • Limitations: The $50,000 income threshold corresponds to the 76th US income quantile in 1994, but the 88th Black quantile and 89th female quantile.
  • Limitations: A logistic regression model reaches 85% overall accuracy, compared with 91.4% on Black instances and 92.7% on female instances, an atypical pattern for historically disadvantaged groups.
  • Threshold sensitivity: The authors vary the income threshold from $6,000 to $70,000 and evaluate accuracy, demographic parity, equality of opportunity, and three intervention classes.

3 New datasets for algorithmic fairness

The paper builds a broader ecosystem of Census-derived fairness datasets and prediction tasks, extending beyond UCI Adult across domains, years, and US states. These resources support empirical study of geographic and temporal variation while remaining benchmarks for algorithmic fairness rather than substantive inequality research.

  • Contribution: The authors use US Census data to create new machine-learning tasks for empirical algorithmic-fairness research.The work draws on the Census data ecosystem to establish stronger empirical evaluation practices.
  • Available prediction tasks: Each task includes age, race, and sex features and can be instantiated across 51 jurisdictions and five years, yielding 255 datasets per task.The jurisdictions are all fifty states plus Puerto Rico, and the years are 2014–2018.
  • Implementation: The folktables package provides an API for accessing the tasks and constructing additional prediction tasks from ACS PUMS data.The authors explicitly encourage exploration beyond the introduced tasks.
  • Scope and limitations: The paper distinguishes Census-based fairness benchmarks from substantive social-science investigations of inequality.Its contribution concerns empirical evaluation of fairness methods, criteria, causal scenarios, and transfer across contexts, not direct substantive claims about inequality.

4 A tour of empirical observations

The experiments use geographic and temporal variation to examine fairness interventions, intervention loci, and whether larger or newer datasets mitigate disparities. They find substantial state-level variation and unpredictable geographic transfer, while fairness remains more stable than accuracy over time.

  • Experimental scope: The experiments examine population variation, intervention locus, and whether dataset growth or social progress mitigates disparities.They use state and year variation to study interventions across distributions and under geographic or temporal shift.
  • Variation within the population: Intervention effects vary markedly across states, with LFR reducing demographic parity violation by 10% in some states but nearly zero in others.Post-processing accuracy drops range from less than 1% in some states to about 5% in others.
  • Variation within the population: Training on one state and testing on another generally produces unpredictable accuracy and fairness results.A post-processing method that enforces demographic parity in one state may not satisfy the constraint when deployed in another.
  • Variation within the population: Fairness criteria are more stable over time than predictive accuracy under temporal distribution shift.Models trained on 2014 US-wide ACSIncome data lose accuracy in later years, while demographic parity violation remains essentially constant.
  • Specifying a locus of intervention: State-level interventions improve US-wide accuracy while preserving demographic parity through post-processing or further mitigating violations through LFR.The comparison aggregates separately trained and intervened state models against a US-level strategy.
  • Increased dataset size does not necessarily mitigate observed disparities: Larger datasets and later data do not automatically mitigate tabular error-rate disparities because persistent social inequality remains reflected in the data.The paper notes that unequal base rates constrain simultaneous calibration and equal error rates except at 100% accuracy.

5 Discussion and future directions

The authors frame their findings as a starting point for further evaluation rather than settled conclusions. They identify open questions about fairness interventions, Bayes-optimality, and geographic variation.

  • The authors evaluated only some methods and call for comprehensive evaluation of existing methods across all datasets.
  • They conjecture that threshold adjustment may outperform demographic-parity or error-rate-parity methods when paired with a nearly Bayes-optimal classifier.
  • The discussion distinguishes settings where a nearly Bayes-optimal classifier is known from those where it is not.
  • The datasets remain unexplored test cases for causal inference methods, including invariant risk minimization across geographic contexts.
  • The reconstruction includes exceptions for native-country matching and excludes fnlwgt because no exact match was identified and it is a population-representativeness weight.

A.2 Varying the income threshold experiments

The experiment reconstructs IPUMS Adult, varies the income-label threshold, and compares unconstrained and fairness-constrained classifiers. The reported comparison examines accuracy and fairness as thresholds change.

  • 49,531 reconstructed examples are split into 32,094 training and 13,755 test examples, with income thresholds ranging from 6,000 to 72,000.
  • For each threshold, the task predicts whether an individual’s income exceeds that threshold.
  • The unconstrained classifier is a gradient boosted decision tree with exponential loss, five estimators, and maximum depth five.
  • Figure 5 compares ExpGrad and post-processing for equality of opportunity, excluding LFR because it does not enforce that criterion without modification.
  • As the threshold varies, the accuracy drop needed to enforce fairness changes, and the equality-of-opportunity results resemble the demographic-parity experiment.

B New prediction task details

This appendix specifies the variables, filters, and targets used to construct the paper’s prediction tasks. The tasks cover income, health coverage, mobility, and commute time, with Census-coded features and explicit eligibility conditions.

  • The task definitions list each feature’s ACS PUMS variable code, extended description, and permitted value range.
  • The income task predicts whether a working adult’s yearly income exceeds $50,000.
  • The income target labels individuals 1 when PINCP > 50000 and otherwise 0, while the software permits changing the threshold.
  • The public-health-coverage task predicts coverage among low-income individuals not eligible for Medicare, labeling PUBCOV == 1 as positive.
  • The mobility task predicts whether a young adult moved addresses during the previous year, using MIG == 1 as the positive label.
  • The transportation task predicts whether a working adult’s commute exceeds 20 minutes.

B.6 Dataset access and license

The software package supports downloading ACS PUMS data and constructing the paper’s tasks or new tasks. Data access is governed by Census Bureau and IPUMS terms of service.

  • The software package downloads ACS PUMS data and constructs the prediction tasks discussed in the paper or new tasks from ACS PUMS products.
  • ACS PUMS data use is governed by the US Census Bureau’s terms of service.
  • The IPUMS Adult reconstruction is governed by IPUMS terms of use.

B.7 Table 1 experiment details

The experiments use standardized 2018 US-Wide ACS PUMS data across five prediction tasks, with state-level sampling and fixed training/testing splits.

  • Each task uses 2018 US-Wide ACS PUMS data, with at most 100,000 examples sampled from each state.
  • The data are randomly split into 80% training and 20% test sets, and all features are standardized to zero mean and unit variance.
  • Constant Predictor, logistic regression, and gradient boosted decision tree classifiers provide the stated baseline models.

C Tour of empirical observations: missing experimental details

The empirical studies use a common GBM-centered setup, with task-specific state or US-wide datasets, confidence intervals, and a CPU cluster environment.

  • All experiments use an unconstrained gradient boosted decision tree because it trained quickly and outperformed logistic regression and random forests in the unconstrained setting.Experiments with other base models produced qualitatively similar results.
  • The GBM uses exponential loss, five estimators, maximum depth five, and default values for other hyperparameters.The hyperparameters were selected through a small grid search.
  • Except for temporal-shift experiments, ACSIncome analyses use 2018 data with state-level 80/20 train-test splits and a US-Wide dataset combining all states and Puerto Rico.
  • Accuracy is reported with 95% Clopper-Pearson intervals, while demographic-parity and equality-of-opportunity violations use 95% Newcombe intervals.
  • Experiments run on CPUs in a cluster with 24 Intel Xeon E7 CPUs and 300 GB of RAM.

D Additional experiments

Additional experiments extend the core analyses across five prediction tasks and compare fairness interventions across state distributions and geographic transfer settings.

  • The additional-task experiments replicate Section 4 with identical models, hyperparameters, and other experimental details.
  • The intervention comparison evaluates unconstrained GBMs against LFR, fair reductions, and post-processing that adjusts group acceptance thresholds.The comparison is conducted for fairness-criterion violations and accuracy.
  • Figure 6 represents each state’s intervention effect as an arrow from the unconstrained model’s accuracy and equality-of-opportunity violation to the intervened model’s point.
  • The experiments also evaluate models on same-state in-distribution data and on out-of-distribution data from the other 49 states.This geographic-transfer analysis is applied across ACSIncome and additional ACS tasks.

D.3 Temporal distribution shift

Temporal-shift experiments evaluate models trained on 2014 data against later years, while accompanying geographic-transfer figures examine fairness and accuracy across states and tasks.

  • D.3 Temporal distribution shift: Models trained on 2014 US-wide ACSIncome data are evaluated on test sets from 2014–2018 under equality-of-opportunity interventions.
  • D.3 Temporal distribution shift: State-to-state transfer produces unpredictable predictive-accuracy and fairness-criterion results for ACSIncome, ACSPublicCoverage, ACSEmployment, ACSMobility, and ACSTravelTime.
  • D.3 Temporal distribution shift: For temporal experiments, accuracy generally declines over time while equality-of-opportunity violations remain essentially constant.
  • D.3 Temporal distribution shift: The corresponding demographic-parity experiments cover four additional ACS tasks using models trained in 2014 and evaluated in subsequent years.
  • D.3 Temporal distribution shift: The paper’s datasheet covers both the introduced prediction tasks and their underlying US Census data sources, with updated documentation maintained in the project repository.

E.2 Composition

The datasets represent sampled individuals from the US population using feature-based records derived from Census data. Their construction includes documented preprocessing, missing values, and sampling weights that affect interpretation and use.

  • Each instance represents an individual and consists of features; the source ACS data also include household-level information and individual–household relationships.
  • The IPUMS Adult reconstruction contains 49,531 rows, while the ACS dataset sizes are reported in Table 1.
  • The datasets are samples of the US population, and per-instance weights are needed when representing the US population.
  • IPUMS Adult uses income as its label, with continuous values available alongside the binarized values; ACS task labels are documented separately.
  • Some features contain missing values, and the datasets identify subpopulations through features such as age, gender, and race.
  • ACS data include top-coded variables and imputed missing values, while task construction commonly maps NaNs to −1.

E.5 Uses

The dataset suite defines five ACS prediction tasks spanning income, health coverage, mobility, employment, and commuting time, while supporting new task definitions through an API. Its use is bounded by sampling noise, subgroup sparsity, privacy concerns, and source-specific terms of use.

  • The five ACS tasks predict income, public health coverage, residential mobility, employment, and travel time exceeding 20 minutes.
  • New tasks can use different feature subsets or target variables, producing diverse Bayes error rates and subgroup base-rate disparities for benchmarking.
  • State-level trends may be noisy because some states contain few representatives of particular subgroups, making generalization from those individuals highly inaccurate.
  • Benchmarking fairness algorithms with few subgroup representatives may create an appearance of fairness without substantive merit.
  • The datasets should not be used for re-identification or as the sole basis for substantive investigations into inequality and demographic shifts.
Loading 2108.04884v3…