Machine Unlearning for Federated Socially Fair k-Means Clustering

Aravind Reddy; Deeraj S K; Mohd Wasif Raza; Saurav Prakash

cs.LGOct 9, 2026

Machine Unlearning for Federated Socially Fair k-Means Clustering

Authors: Aravind Reddy, Deeraj S K, Mohd Wasif Raza, Saurav Prakash

Organizations: Centre for Responsible AI, Wadhwani School of Data Science & AI, Indian Institute of Technology Madras · Department of Artificial Intelligence, Sardar Vallabhbhai National Institute of Technology, Surat, India · Department of Electrical Engineering, Indian Institute of Technology Madras

Abstract

Federated clustering can support applications such as patient stratification across hospitals and customer segmentation across distributed financial datasets. Responsible AI calls for fair representation of demographic groups and support for data deletion, motivated in part by rights such as the GDPR's right to erasure. These goals interact: deleting records can change the balance between groups and alter the clustering of the remaining population. We study exact machine unlearning for federated socially fair $k$ -means, whose objective is to minimize the largest group-average squared distance to the nearest center. Our approach keeps separate local summaries for each group, so summaries at unaffected institutions remain valid as global fairness weights change. For two groups, we prove an initialization approximation guarantee independent of client group proportions under ideal sampling, and we show that same-seed replay reproduces retraining exactly. Reuse pays off most when a whole client withdraws: across three datasets, replay is $35.8\times$ faster than retraining for the initialization-only model and remains $2.4\times$ and $1.7\times$ faster with one and two fair-refinement rounds, which improve worst-group cost. For individual record deletions, gains are modest (about $1.06\times$ ), and certifying cached assignments can cost more than it saves. These results map where exact unlearning can profitably reuse computation. Code and reproducibility materials are available at https://github.com/arareddy/mufsc.

Figures & tables

Suite	Method	Total algorithm	Total E2E	Algorithm losses	$S/N$ (%)
P2	Direct	1.061	1.049	2/175	0.00
P2	Basic	1.006	0.996	68/175	0.76
P2	Runner-up	1.005	0.995	67/175	0.76
P3	Direct	1.056	1.043	15/95	0.00
P3	Basic	0.938	0.930	74/95	1.35
P3	Runner-up	0.940	0.931	76/95	1.35

Table 1: Does replay save time? Total ratios divide summed fresh times by summed replay times over all trials; values above one favor replay. Algorithm time measures the call; E2E also includes input materialization, checkpoint loading and equal output writing. Losses count algorithm-time comparisons with fresh. Useful coverage is pooled

S/N

, the share of survivor point-rounds whose assignments certification actually saved. All 525 P2 and 285 P3 replay comparisons match exactly.

	Compact $T=0$	Direct $T=1$	Direct $T=2$	$\Phi(T=0)$	$\Phi(T=2)$
Adult	26.96	2.053	1.544	117.957	95.998
Bank	115.53	2.646	1.828	45.898	32.634
Credit	25.03	2.647	1.856	20.124	13.009

Table 2: Whole-client deletion at

k=10

: one preselected client per dataset, five seeds, on the same retained populations. Ratios divide total fresh request time by total replay request time at the same

T

(covering resident preparation, the call and output writing); values above one favor replay. Compact

T=0

uses a separate implementation and three repeats; canonical Direct

T=1,2

uses four repeats. Quality is mean worst-group cost over five seeds, computed within seed before averaging.

Deletion population	Dataset	Worst-group cost	Population-average cost
Record	Adult	$-4.75$	$+4.72$
Record	Bank	$-3.15$	$+0.55$
Record	Credit	$-5.07$	$+0.71$
Client	Adult	$-3.92$	$+4.47$
Client	Bank	$-4.24$	$+1.87$
Client	Credit	$-4.93$	$+0.53$

Table 3: What does fair refinement improve? At

T=2

, fair-minus-Lloyd percentage differences

100(c_{\rm fair}-c_{\rm Lloyd})/c_{\rm Lloyd}

are computed within seed, then averaged over five seeds. Survivors and fair-initialized centers are matched. Negative favors fair for that cost. This compares update rules, not optimized ordinary

k

-means.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Mean $G/G^{*}$	SE	Mean $\Phi/\Phi^{*}$
Merged-1k	63.5310	1.1028	32.3935
Merged-2k	1.9034	0.1518	1.4593
SplitGroup	2.2375	0.1776	1.6270
Central	1.7749	0.1406	1.3948

Appendix

Table 4: P1 at

\kappa=128

, 10,000 seed indices per method. SE describes across-seed variation under the declared deterministic PRNG protocol, with the ideal-bit expectation a separate theorem.

Figure 1: P1: does locally balanced merging lose quality as the local/global mismatch

\kappa

grows? Initialization quality on the lower-bound construction of Appendix C , with 10,000 seed indices per method and heterogeneity value. Dashed curves are ideal analytical expectations; error bars are descriptive normal intervals from across-seed standard errors. The increasing Merged-1k cost concerns the specified two-location construction, not every merged summary.

Dataset	Records	Clients	$d$	Group 0	Group 1	$\kappa$
Adult	30162	100	103	9782	20380	1.317781
Bank	45211	20	42	17997	27214	1.008094
Credit	30000	100	23	5385	24615	2.416867

Appendix

Table 5: Source-cohort manifest, p124-unique-source-v2. All listed source clients are unique; source hashes match the audit and the full-data partition is byte-verified. No client is pure or at least 95% single-group.

Method	Mean	Mean 95% CI	Median	Wins/fresh	Abandons
Direct	1.066	[1.065, 1.068]	1.067	173/175	N/A
Basic	1.025	[1.004, 1.062]	1.008	107/175	174/175
Runner-up	1.024	[1.004, 1.061]	1.007	108/175	174/175

Appendix

Table 6: P2, all 175 trial units. Ratios use output-matched serial elapsed time. Intervals are 5,000 fixed-

k

seed-block bootstrap replicates, preserving shared seeds across datasets. Every replay trial saves exact indexed trajectories and raw statistics.

Method	Elapsed range	Model mean	E2E mean
Direct	0.960–1.138	1.002	1.052
Basic	0.938–4.222	0.985	1.010
Runner-up	0.934–4.187	0.986	1.009

Appendix

Table 7: P2 cost detail, ratios of fresh/method time. End-to-end includes recorded input/checkpoint/materialization/witness-output cost; modeled time is compute only.

Method	$N$	$A$	$P$	$S$	$J$	$P/A$
Direct	89086500	0	0	0	0	N/A
Basic	89086500	6571424	677014	677014	5893934	10.3024%
Runner-up	89086500	6571424	677018	677014	5893930	10.3025%

Appendix

Table 8: P2 point-round counters pooled over 175 trials.

N,A,P,S,J

are defined in Section I.2 ;

P/A

includes discarded passes, while useful coverage is

S/N

. Direct replay attempts no certificates.

$n$	$C$	Client	$n_{A}$	$n_{B}$	$\kappa$	Client size	P/NP
100,000	10	State	48,728	51,272	1.105216	1446–41266	0/0
100,000	25	State	49,022	50,978	1.127072	392–23054	0/0
100,000	50	State	49,033	50,967	1.229392	179–11808	0/0
100,000	100	PUMA	48,925	51,075	1.374730	383–2254	0/0
100,000	250	PUMA	49,293	50,707	1.523855	150–884	0/0
250,000	50	State	122,606	127,394	1.183868	447–29521	0/0

Appendix

Table 9: ACS populations. PUMA identities are state-qualified. Group counts, client-size ranges, and exact count-derived heterogeneity are shown. P/NP counts pure and at least 95% single-group clients; these are population descriptors, not performance explanations.

ID	$n$	$C$	$k$	$T$	$r$	$r/n$
0	100,000	50	4	5	100	0.001
1	250,000	50	4	5	250	0.001
2	500,000	50	4	5	500	0.001
3	1,000,000	50	4	5	1,000	0.001
4	100,000	10	4	5	100	0.001
5	100,000	25	4	5	100	0.001

Appendix

Table 10: Frozen P3 one-factor design. Every row uses seed indices 0–4 and all four methods;

d=16,b=12,L=6,T_{a}=0,\gamma=0

, abandonment threshold

0.5

. Designs 15 and 16 are the prespecified pilots. Repeated defaults across one-factor sweeps form a union of 19 configurations, not a factorial experiment.

Method	Mean E2E	Total E2E	Losses	Mean $P/A$	Pool $P/A$	Def.
Fresh	1.000	1.000	0/95	N/A	N/A	0/95
Direct	1.078	1.043	23/95	N/A	N/A	0/95
Basic	1.252	0.930	78/95	3.16	8.38	95/95
Runner-up	1.228	0.931	79/95	3.16	8.38	95/95

Appendix

Table 11: Additional descriptive P3 suite summaries. End-to-end (E2E) means and total-time ratios are distinct. Losses are paired elapsed E2E losses against fresh. Acceptance percentages exclude undefined trial ratios from the mean; Def. gives the count. No IID interval is attached to this varied suite.

ID	Direct	Basic	Runner-up
0	1.079 [1.072, 1.086] 1.079 [1.071, 1.086]	0.918 [0.913, 0.923] 0.918 [0.913, 0.923]	0.919 [0.913, 0.925] 0.919 [0.913, 0.925]
1	1.026 [1.016, 1.035] 1.026 [1.016, 1.034]	0.882 [0.872, 0.890] 0.882 [0.872, 0.890]	0.876 [0.866, 0.884] 0.876 [0.866, 0.884]
2	1.020 [1.007, 1.035] 1.020 [1.007, 1.035]	0.878 [0.867, 0.889] 0.878 [0.867, 0.889]	0.872 [0.861, 0.887] 0.872 [0.861, 0.887]
3	1.011 [0.996, 1.029] 1.011 [0.996, 1.029]	0.871 [0.865, 0.877] 0.871 [0.865, 0.877]	0.872 [0.861, 0.882] 0.872 [0.861, 0.882]
4	0.999 [0.984, 1.013] 0.999 [0.984, 1.013]	0.859 [0.844, 0.877] 0.859 [0.844, 0.877]	0.864 [0.853, 0.873] 0.864 [0.853, 0.873]
5	1.036 [1.025, 1.049] 1.036 [1.025, 1.049]	0.898 [0.879, 0.915] 0.898 [0.879, 0.915]	0.874 [0.853, 0.892] 0.874 [0.853, 0.892]

Appendix

Table 12: P3 algorithm elapsed ratios Fresh/method for all 19 designs. Each cell’s first line is the mean paired ratio [95% percentile interval]; its second line is the ratio of total times [interval]. The paired bootstrap resamples five seed indices within each fixed design, 5,000 draws, seed 20260906. Intervals are descriptive small-sample uncertainty, not deployment reliability. Ratios above one favor replay.

Design/seed	Direct	Basic	Runner-up	Fallback
15/0	1.374 / 1.342	17.331 / 11.488	15.544 / 10.731	Never
15/1	1.355 / 1.324	16.959 / 11.377	15.191 / 10.591	Never
15/2	1.386 / 1.353	1.145 / 1.126	1.144 / 1.125	0
15/3	1.386 / 1.353	1.151 / 1.130	1.138 / 1.119	0
15/4	1.337 / 1.308	14.438 / 10.241	13.026 / 9.504	Never
16/0	1.258 / 1.232	1.070 / 1.055	1.065 / 1.049	0

Appendix

Table 13: When does certification pay off? All small-deletion trials, as algorithm / E2E paired fresh-to-method ratios. Certification reaches large ratios only in the three trials that never fall back; in the other seven it beats fresh training here but trails direct replay. Values above one favor replay. Designs 15 and 16 have

r=1

and

r=10

; other settings are in Table 10 . Fallback is the first zero-based round, identical for both certificates. Direct has no fallback. Both certificates’ attempted acceptance and useful coverage round to 100% in the three never-fallback trials and are zero in the other seven; exact unsaved counts for the successes appear in Table 15 . Full per-method seconds and unrounded records remain in the artifact.

Method	$N$	$A$	$P$	$S$	$J$
Fresh	110,998,475	0	0	0	0
Direct	110,998,475	0	0	0	0
Basic	110,998,475	17,905,183	1,499,968	1,499,968	16,405,198
Runner-up	110,998,475	17,905,183	1,499,968	1,499,968	16,405,198

Appendix

Table 14: P3 raw point-round totals over the 95-trial suite.

N

counts eligible assignments,

A

attempts,

P

attempted passes,

S

effective saved assignments, and

J

duplicate assignments. Non-certifying methods have no attempts and hence undefined

P/A

, not zero acceptance.

Design/seed	$r$	Changed labels/ $N_{0}$	$\Delta_{\rm idx}$	$\Delta_{\rm match}$	$F_{0}$	Server draw
15/0	1	0/99999	0.000	0.000	0	none
15/1	1	0/99999	0.000	0.000	0	none
15/2	1	43870/99999	4.302	3.867	99999	0
15/3	1	31280/99999	6.471	6.264	99999	2
15/4	1	0/99999	0.000	0.000	0	none
16/0	10	60120/99990	8.880	7.292	99990	0

Appendix

Table 15: Seven prescribed ACS diagnostic cases: why did certification fail or succeed? The top panel is round zero:

N_{0}=100000-r

, indexed label changes, maximum indexed movement and diagnostic bottleneck movement (rounded to three decimals), certificate failures

F_{0}

, and first server draw with changed selected coordinates. Both implemented and independently resolved ideal Basic/Runner-up failure counts equal the displayed

F_{0}

; every numerical-only and unresolved count is zero. Matching never defines exactness. The lower panel reuses full five-round accounting,

N=5N_{0}

: attempts

A

, passes

P

, effective coverage

S/N

, duplicated assignments

J

, and first fallback round. Counts coincide for the two certificate methods in these seven cases. Draw/round indices start at zero; “none” means no such event. The three successful trials leave

3+6+8=17

point-rounds unsaved despite percentages that round to

100\%

Dataset	Method	$Q$	Base IDs	Extra IDs / slots	Packet bytes $0/1$
Adult	Merged-1k	8.81718	5	26/800	697000/697000
Adult	Merged-2k	8.79206	14	66/1600	1369000/1369000
Adult	SplitGroup	8.77216	8	70/1600	1377200/1377200
Bank	Merged-1k	6.95398	0	0/160	61320/61320
Bank	Merged-2k	6.98036	0	2/320	117640/117640
Bank	SplitGroup	7.00838	0	0/320	119280/119280

Appendix

Table 16: P4 at

\gamma=1

Q

uses actual represented anchors with global group-normalized weights. Base IDs are duplicates before snapping; extra IDs show additional collisions over all anchor slots. Packet bytes compare

\gamma=0/1

and include the implemented metadata; they are not network traffic.

Figure 2: P4: how does the anchor grid width

\gamma

affect quality and anchor residuals? All 57 settings on the unique-source cohorts, seed 42. Top: final represented-data group-maximum cost. Bottom: unweighted mean squared residual over anchor slots, not weighted

Q

. Dashed lines are centralized quality references. No variability across training seeds is estimated.

Dataset	$T$	$\Phi$ : mean $\pm$ SD	Direct	Basic	Runner-up
Adult	0	119.594 $\pm$ 3.048	4.119/3.304	4.101/3.284	4.110/3.296
Adult	1	97.941 $\pm$ 1.184	1.741/1.647	1.050/1.024	1.044/1.019
Adult	2	95.931 $\pm$ 1.391	1.421/1.374	1.033/1.014	1.031/1.012
Bank	0	47.959 $\pm$ 0.757	1.390/1.345	1.394/1.349	1.394/1.348
Bank	1	34.016 $\pm$ 1.106	1.228/1.203	0.889/0.881	0.890/0.882
Bank	2	33.010 $\pm$ 0.929	1.161/1.142	0.917/0.908	0.914/0.905

Appendix

Table 17: Record-deletion results, five seeds and four timing repetitions. Each ratio cell is algorithm / file-I/O-inclusive E2E fresh-to-replay total time.

T

counts refinement rounds. Quality is identical across methods at a matched seed and budget.

Figure 3: Record deletion across

T=0,1,2

: five-seed means of four-repeat median algorithm times, with one seed SD of

\Phi

. Fresh and Direct have identical matched-budget quality. Connecting measured budgets does not establish a Pareto frontier or a cross-budget unlearning ratio.

Dataset	Seed	Fresh ms	Direct ms	Fast ms	F/fast	$\Phi_{0}$	$\Phi_{2}$
Adult	10000	244.925	17.138	9.019	27.16	112.512	95.182
Adult	10001	243.347	17.308	9.070	26.83	117.132	95.807
Adult	10002	242.373	17.141	9.162	26.46	117.421	97.242
Adult	10003	250.312	17.690	9.403	26.62	119.269	98.305
Adult	10004	248.264	17.986	8.950	27.74	123.448	93.452
Bank	10000	221.181	6.316	1.973	112.12	44.867	32.578

Appendix

Table 18: Whole-client case means: three timing repetitions per method, with a single preselected client per dataset. All speed ratios use fresh

T=0

;

\Phi_{2}

is solely a quality reference.

Dataset	Train	Compact	Load	Full ser.	Small ser.	Full MiB	Small MiB
Adult	251.805	1.178	44.902	5.627	1.202	52.481	3.196
Bank	231.292	0.173	30.118	2.455	0.184	29.734	0.267
Credit	167.754	0.669	17.164	2.887	0.869	11.929	0.754

Appendix

Table 19: Whole-client setup means across five seeds. Times are milliseconds; protocol-5 serialized sizes are measured in memory. Serialization is accounted separately, not as durable per-request state persistence.

Dataset	$T$	Fresh/Direct	Fresh/Basic	Fresh/Runner-up
Adult	1	2.053	1.164	1.159
Bank	1	2.646	1.455	1.464
Credit	1	2.647	1.425	1.407
All	1	2.357	1.309	1.305
Adult	2	1.544	1.107	1.103
Bank	2	1.828	1.287	1.286

Appendix

Table 20: Positive-budget whole-client replay. Ratios divide total resident-request times over five seeds and four repeats per dataset/budget; “All” divides pooled seconds. Each replay method wins all 15 seed/dataset cases per budget against fresh.

Dataset	Client ID	Removed	Removed group 0/1	Retained group 0/1	Anchors
Adult	0	300	97/203	9,685/20,177	1,980
Adult	99	462	179/283	9,603/20,097	1,980
Bank	0	2,259	899/1,360	17,098/25,854	380
Bank	19	2,290	916/1,374	17,081/25,840	380
Credit	0	299	53/246	5,332/24,369	1,980
Credit	99	399	138/261	5,247/24,354	1,980

Appendix

Table 21: Metadata-selected client coverage. Requests independently start from the original checkpoint. Both groups survive each removal; within each dataset both requests retain the same number of anchor slots.

	Matched resident output			Including buffered state round trip
Dataset	Fresh ms	Compact ms	Ratio	Fresh ms	Compact ms	Ratio
Adult	713.326	26.608	26.81	714.219	44.826	15.93
Bank	612.280	5.763	106.25	613.038	8.346	73.45
Credit	483.370	19.331	25.01	484.338	28.164	17.20

Appendix

Table 22: Cumulative three-withdrawal timings at matched

T=0

, averaged over five seeds and three repeats. Operational time adds each method’s necessary buffered state write/reload to its matched resident-output time; it is not durable transaction latency.

Figure 4: Ratios of cumulative fresh/compact times through each withdrawal. Shading spans the five paired seed ratios, not a confidence interval. Blue uses matched resident model outputs; orange also charges method-specific buffered state persistence/reload. Steps and timing repetitions are correlated; dataset scales differ.

Population	Seed	Group 1 before	Group 1 after	Change (%)	Population change
Record	10000	96.192670	97.326494	+1.1787	+0.039075
Record	10002	96.224892	96.477361	+0.2624	$-0.048872$
Record	10004	94.487752	96.450765	+2.0775	+0.593167
Client	10002	96.002141	96.957289	+0.9949	$-0.179302$
Client	10003	94.642231	95.233313	+0.6245	$-0.155996$

Appendix

Table 23: Every adverse fair group-cost transition: Adult,

T=1

T=2

. Worst-group cost decreases in every row. Population-average changes are signed absolute differences.

Figure 5: Record-deletion quality on fixed represented survivors: separate group costs (top), worst-group and population-average costs (bottom). Thin lines retain five paired seeds; bold markers and bars show mean and one sample SD. Worst-group maxima and population count weighting are computed within seed. Adult group 1 rises in three second-round transitions. Dataset axes differ and have nonzero origins.

Figure 6: Whole-client-deletion quality on the separate fixed client-0 survivor populations, with the same metric and seed conventions as Figure 5 . Adult group 1 rises in two second-round transitions. Positive-budget quality references are not denominators for zero-budget unlearning speedups.

Figure 7: Matched-start fair versus Lloyd refinement on separate record- and whole-client-deletion populations. Each seed’s difference is

100(c_{\rm fair}-c_{\rm Lloyd})/c_{\rm Lloyd}

; negative favors fair for that cost. Thin lines show five seeds, bold markers their mean, bars one sample SD. Shared initialization gives zero differences at

T=0

; axes differ by dataset.

Population	Dataset	Worst group	Population	Worst group	Population
		$T=1$		$T=2$
Record	Adult	$-3.57$	+4.80	$-4.75$	+4.72
Record	Bank	$-2.54$	+0.57	$-3.15$	+0.55
Record	Credit	$-3.26$	+0.59	$-5.07$	+0.71
Client	Adult	$-2.93$	+4.02	$-3.92$	+4.47
Client	Bank	$-4.25$	+1.62	$-4.24$	+1.87

Appendix

Table 24: Fair-minus-Lloyd mean paired percentage differences:

100(c_{\rm fair}-c_{\rm Lloyd})/c_{\rm Lloyd}

within seed, then averaged. Negative favors fair for the indicated cost. Each row has five seeds;

T=0

is exactly tied.

Clustering

Clustering groups data points by shared structure, developing methods to discover, validate, and interpret groups across diverse data and constraints.

See all 342 papers

Jun 3, 2026

UniFair: A unified fair clustering approach based on separation and compactness

Clustering is increasingly used to support high-impact decisions, yet standard objectives such as k-means can produce clusterings that treat demographic groups unequally. Existing fair clustering methods typically optimize a single notion of fairness and often overlook how clustering costs interact with the geometry of the induced decision boundaries. We propose UniFair, a unified framework that jointly optimizes separation fairness and social fairness. Separation fairness encourages protected groups to lie farther from the induced decision boundaries, while social fairness reduces disparities in within-cluster distortion by penalizing group-wise clustering costs. We develop gradient-based optimization procedures for separation-fair and unified k-means objectives, and extend them to deep clustering by enforcing the same criteria in the latent space of an autoencoder. Experiments on tabular and image datasets show that UniFair reduces both boundary-related and cost-based group disparities with only a modest increase in clustering loss.

Antonia Karra, Vasiliki Papanikou, Georgios Vardakas +2

University of Ioannina, Greece · Archimedes/Athena Research Center, Greece

Clustering Algorithmic Fairness Deep Clustering Group Fairness

Ask PDF

May 23, 2026

Rethinking Federated Unlearning via the Lens of Memorization

Federated learning (FL) increasingly needs machine unlearning to comply with privacy regulations. However, existing federated unlearning approaches may overlook the overlapping information between the unlearning and remaining data, leading to ineffective unlearning and unfairness between clients. In this work, we revisit federated unlearning through the lens of memorization. We argue that unlearning should mainly remove the unique memorized information attributable to the data to be forgotten, while preserving overlapping patterns that are also supported by the remaining data. Specifically, we propose Grouped Memorization Evaluation, an example-level metric that separates memorized knowledge from overlapping knowledge. Building on this metric, we introduce Federated Memorization Pruning (FedMemPrune), a pruning-based unlearning approach that resets redundant parameters responsible for memorization. Extensive experiments show that FedMemPrune closely matches retraining-based unlearning baselines while more effectively eliminating memorization than existing federated unlearning algorithms, yielding strong unlearning performance without sacrificing the utility of retained knowledge.

Jiaheng Wei, Yanjun Zhang, He Zhang +5

Royal Melbourne Institute of Technology · Griffith University · Swinburne University of Technology

Federated Unlearning Neural Network Memorization Machine Unlearning Federated Learning

Ask PDF

Apr 27, 2023

Proportionally Representative Clustering

In recent years, there has been a surge in effort to formalize notions of fairness in machine learning. We focus on centroid clustering--one of the fundamental tasks in unsupervised machine learning. We propose a new axiom ``proportionally representative fairness'' (PRF) that is designed for clustering problems where the selection of centroids reflects the distribution of data points and how tightly they are clustered together. Our fairness concept is not satisfied by existing fair clustering algorithms. We design efficient algorithms to achieve PRF both for unconstrained and discrete clustering problems. Our algorithm for the unconstrained setting is also the first known polynomial-time approximation algorithm for the well-studied Proportional Fairness (PF) axiom. Our algorithm for the discrete setting also matches the best known approximation factor for PF.

Haris Aziz, Barton E. Lee, Sean Morota Chu +1

UNSW Sydney · ETH Zürich · Northwestern University

Clustering Algorithmic Fairness Group Fairness Approximation Algorithms

Ask PDF

Abstract

Clustering

Explore similar work

UniFair: A unified fair clustering approach based on separation and compactness

Rethinking Federated Unlearning via the Lens of Memorization

Proportionally Representative Clustering