Selective Data Expansion for Model Performance

Jahid Hasan; Romila Pradhan

cs.LGDec 4, 2024

Selective Data Expansion for Model Performance

Authors: Jahid Hasan, Romila Pradhan

Organizations: Purdue University, West Lafayette, USA

Abstract

Machine learning systems are increasingly being used in critical decision making such as healthcare, finance, and criminal justice. Concerns around their fairness have resulted in several bias mitigation techniques that emphasize the need for high-quality data to ensure fairer decisions. However, the role of earlier stages of machine learning pipelines in mitigating model bias has not been explored well. In this paper, we focus on the task of acquiring additional labeled data points for training the downstream machine learning model to rapidly improve its fairness. Since not all data points in a data pool are equally beneficial to the task of fairness, we generate an ordering in which data points should be acquired. We present DataSift, a data acquisition framework based on the idea of data valuation that relies on partitioning and multi-armed bandits to determine the most valuable data points to acquire. Over several iterations, DataSift selects a partition and randomly samples a batch of data points from the selected partition, evaluates the benefit of acquiring the batch on model fairness, and updates the utility of partitions depending on the benefit. To further improve the effectiveness and efficiency of evaluating batches, we leverage influence functions that estimate the effect of acquiring a batch without retraining the model. We empirically evaluate DataSift on several real-world and synthetic datasets and show that the fairness of a machine learning model can be significantly improved even while acquiring a few data points.

Figures & tables

Figure 1. Impact of different data selection strategies on model fairness (statistical parity) and classification accuracy. Lower magnitude of fairness indicates a less biased and more equitable model overall.

Figure 2. Overview of DataSift for selective data expansion aimed at improving model fairness. The blue dashed box highlights DataSift with a reinforcement learning–based multi-armed bandit (MAB) framework for selective data expansion, while the red dashed component incorporates data valuation to prioritize informative and fairness-enhancing samples.

Evaluation	$\Delta\mathcal{F}$	$U_{1}$	$\Delta\mathcal{F}$	$U_{2}$	$\Delta\mathcal{F}$	$U_{3}$	$\Delta\mathcal{F}$	$U_{4}$	Decision
	$C_{1}$		$C_{2}$		$C_{3}$		$C_{4}$
1	-0.0118 $\downarrow$	-0.0118		-0.0073		-0.0068		-0.0050	$\Delta\mathcal{F}\downarrow$ –Discard batch, Next Cluster $\rightarrow C_{4}$
2		0.0393		0.0382		0.0373	0.0366 $\uparrow$	0.0991	$\Delta\mathcal{F}\uparrow$ –Retain batch, Next Cluster $\rightarrow C_{4}$
3		0.0595		0.0537		0.0510	0.0043 $\uparrow$	0.0976	$\Delta\mathcal{F}\uparrow$ –Retain batch, Next Cluster $\rightarrow C_{4}$
4		0.0800		0.0712		0.0673	0.0110 $\uparrow$	0.0800	$\Delta\mathcal{F}\uparrow$ –Retain batch, Next Cluster $\rightarrow C_{1}$
5	0.0050 $\uparrow$	0.1257		0.1871		0.1819		0.0999	$\Delta\mathcal{F}\uparrow$ –Retain batch, Next Cluster $\rightarrow C_{2}$

Table 1. Illustrative example of cluster and batch selection decisions of DataSift on ACSIncome with

|\mathcal{D}|=50

and batch size

2

\Delta\mathcal{F}

is the round-wise fairness change after adding a batch;

U

is the UCB score. Dark green indicates improved fairness after adding the selected batch; red indicates a decline. The bold

U

value indicates the cluster selected next.

Dataset	Adult Income	Credit	ACSPublicHealth	ACSMobility	ACSEmployment	ACSIncome
Size	$45,222$	$150,000$	$138,554$	$80,329$	$378,817$	$250,847$
No. of features	8	10	19	22	16	11
Protected Group ( $\mathbf{S}=0$ )	Female	Age < 35	Female	African-American	African-American	Female
Population of ( $\mathbf{S}=0$ ) (%)	33	12.8	55.9	5.2	4.8	47.2
Positive Labels in ( $\mathbf{S}=0$ ) (%)	11	11	35	73	39	34
Predictive Task	income> $50K?	serious delay in 2 years?	has public insurance coverage?	moved address last year?	employed?	income> $50K?

Table 2. Summary of datasets.

Panel of Figure 3 — Figure 3. Comparing DataSift with baselines to highlight the effectiveness in achieving fairness. The green shaded area represents the defined zone of unfairness.

Panel of Figure 4 — Figure 4. Evaluation of DataSift over different fairness metrics ( ACSIncome dataset, magenta dashed line: $\mathcal{F}_{{\mathbf{D}}_{train}\cup\mathcal{D}}$ ).

Panel of Figure 5 — Figure 5. DataSift ’s evaluation across parametric models.

Dataset	Methods
	Initial	Random	Entropy	AutoData	SliceTuner	DataSift
AdultIncome	0.80	0.78	0.77	0.71	0.80	0.78
ACSIncome	0.70	0.78	0.77	0.78	0.77	0.77
Employment	0.69	0.75	0.75	0.74	0.74	0.73
PublicHealth	0.65	0.69	0.68	0.68	0.68	0.68
Mobility	0.77	0.77	0.77	0.77	0.75	0.75

Table 3. Accuracy of Logistic Regression models across datasets trained after adding selectively acquired data points from the corresponding pool as determined by each method. (Bold: highest, Underlined: second highest)

Panel of Figure 6 — Figure 6. Comparing DataSift with DataSift_M and Inf to highlight the rationale behind DataSift ’s formation. For each dataset, the left panel shows fairness improvement and the right panel shows accuracy. The magenta and green dotted lines indicate fairness and accuracy, respectively, after adding the entire data pool.

Panel of Figure 7 — Figure 7. Effect of Hyper-parameters and design choices (ACSIncome dataset)

Panel of Figure 8 — Figure 8. Effect of Partitioning and Clustering Sensitivity on Fairness Convergence. (DS: DataSift )

Panel of Figure 9 — Figure 9. Scalability on synthetic dataset.

Figure 10. Efficiency on real-world datasets.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Figure 11. DataSift vs. baselines (True Positive Rate).

Panel of Figure 12 — Figure 12. DataSift vs. baselines (Predictive Parity).

Panel of Figure 13 — Figure 13. Varying data pool distributions.

Algorithmic Fairness

Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.

See all 417 papers

FairBED: Incorporating Fairness into Bayesian Experimental Design
Jun 22, 2026Marcel Hedman, Emily Alger, Brieuc Lehmann +2Algorithmic Fairness Bayesian Experimental Design
Department of Statistics, University of Oxford · Department of Statistical Science, University College London
No evaluation without fair representation : Impact of label and selection bias on the evaluation, performance and mitigation of classification models
Mar 10, 2026Magali Legast, Toon Calders, François FoussAlgorithmic Fairness Dataset Bias
Université catholique de Louvain, Institute of Information and Communication Technologies, Electronics and Applied Mathematics, Louvain-la-Neuve, Belgium · Universiteit Antwerpen, Adrem Data Lab, Antwerpen, Belgium · Université catholique de Louvain, Louvain Research Institute in Management and Organizations, Institute of Information and Communication Technologies, Electronics and Applied Mathematics, Mons, Belgium
Data Bias Mitigation under Coverage Constraints & The Price of Fairness
Jun 18, 2026Bruno Scarone, Alfredo Viola, Renée J. MillerDebiased ML Algorithmic Fairness
Khoury College of Computer Sciences, Northeastern University, USA · Casa de Investigadores Científicos La Comarca, Uruguay · Cheriton School of Computer Science, University of Waterloo, Canada

Jun 22, 2026J/K move · Enter open · S save

FairBED: Incorporating Fairness into Bayesian Experimental Design

Marcel Hedman, Emily Alger, Brieuc Lehmann, Chris Holmes+1

Department of Statistics, University of Oxford · Department of Statistical Science, University College London

Algorithmic Fairness Bayesian Experimental Design Dataset Bias

Open paper AskPDF arXiv

Bayesian experimental design (BED) has traditionally focused exclusively on collecting data that is maximally informative, with little consideration of other desirable dataset characteristics, such as fairness. We therefore extend BED to target data that is both informative and fair. To this end, we introduce FairBED, which provides a formalism for the fairness of datasets themselves by measuring how compatible they are with a fair world, in which the target quantity we wish to learn about is unrelated to the sensitive attributes we want to be fair to. We then use this to construct practical fairness-aware BED objectives that maximise expected information gain about the target variables, while penalising information about sensitive attributes and/or their coupling with the target. We further extend FairBED to active learning, and show empirically that models trained on data gathered using FairBED provide improved fairness-accuracy trade-offs compared to randomly acquired data and conventional BED.

Abstract

Algorithmic Fairness

Explore similar work

FairBED: Incorporating Fairness into Bayesian Experimental Design

No evaluation without fair representation : Impact of label and selection bias on the evaluation, performance and mitigation of classification models

Data Bias Mitigation under Coverage Constraints & The Price of Fairness