Nikhil Garg
LightNews LightNews
nkgarg.bsky.social
Nikhil Garg
@nkgarg.bsky.social
2.6K followers 1K following 300 posts
I study algorithms/learning/data applied to democracy/markets/society. Asst. professor at Cornell Tech. https://gargnikhil.com/. Helping building personalized Bluesky research feed: https://bsky.app/profile/paper-feed.bsky.social/feed/preprintdigest
Posts Replies Media Videos
nkgarg.bsky.social
For example, the NYC Dept of Parks and Recreation no longer crowdsources requests to plant trees. Public data illustrates why: requests are highly correlated with neighborhood income, and *negatively* a measure of need, the heat vulnerability index
Screenshot of a paper Figure showing a map and a regression table. The caption is: In NYC, the number of tree planting requests by ZIP Code Tabulation Area in 2015-2024. Controlling for population, requests correlate positively with neighborhood income and negatively with a heat vulnerability index, a proxy for the need for shade. NYC no longer takes requests to plant trees, and instead will develop a schedule that prioritizes the most heat vulnerable areas.
August 11, 2025 at 1:25 PM
nkgarg.bsky.social
New piece, out in the Sigecom Exchanges! It's my first solo-author piece, and the closest thing I've written to being my "manifesto." #econsky #ecsky
arxiv.org/abs/2507.03600
Screenshot of paper abstract, with text: "A core ethos of the Economics and Computation (EconCS) community is that people have complex private preferences and information of which the central planner is unaware, but which an appropriately designed mechanism can uncover to improve collective decisionmaking. This ethos underlies the community’s largest deployed success stories, from stable matching systems to participatory budgeting. I ask: is this choice and information aggregation “worth it”? In particular, I discuss how such systems induce heterogeneous participation: those already relatively advantaged are, empirically, more able to pay time costs and navigate administrative burdens imposed by the mechanisms. I draw on three case studies, including my own work – complex democratic mechanisms, resident crowdsourcing, and school matching. I end with lessons for practice and research, challenging the community to help reduce participation heterogeneity and design and deploy mechanisms that meet a “best of both worlds” north star: use preferences and information from those who choose to participate, but provide a “sufficient” quality of service to those who do not."
August 11, 2025 at 1:25 PM
nkgarg.bsky.social
Map from @thecity.nyc is much more detailed

www.thecity.nyc/2025/06/24/m...
June 25, 2025 at 2:11 AM
nkgarg.bsky.social
This NYT map illustrates the weirdness of looking just at first place votes. UWS is colored for Cuomo even though he is doing (slightly) worse there than city-wide, because Lander got a good chunk of first place votes
June 25, 2025 at 2:10 AM
nkgarg.bsky.social
Nothing sophisticated, but daily active users of the Paper Skygest went from around 900 to 1100 at the time of your post. Those two spikes in end April and early May are when your post took off, I think
June 22, 2025 at 5:56 PM
nkgarg.bsky.social
See very last tweet...I had privately thought it bad form to imply that it was fake without evidence, but the update seems to vindicate him.
May 16, 2025 at 4:40 PM
nkgarg.bsky.social
Good point, it's hard to tell from Figure 1 what the very recent absolute number trend is. @sjgreenwood.bsky.social at some point we should look at this with our data too!
May 15, 2025 at 12:59 PM
nkgarg.bsky.social
We end with recommendations for what the discrimination auditing and related literatures should do, when it needs to impute race/ethnicity. tl;dr: use continuous scores whenever possible! Basically never use argmax/thresholding.
Putting things together, we caution against the use of existing discrete labels, as commonly distributed in commercial voter files, for sensitive applications where such bias would affect results. For prior results that rely on analyzing discretized demographics, we urge considering how undercounting minority groups or geographically skewing their distribution would affect the results. We further answer the question: what should a researcher or practitioner do to prevent discretization-induced bias? When possible, use the continuous scores instead of single discrete labels, as in (2, 3, 12, 35–38). For example, some auditing tasks can be conducted with sufficiently calibrated scores. In SI Appendix C.1, we show how using continuous scores recovers self-identified results in downstream analyses of voter turnout by race and income, while discretizing leads to substantial bias. However, for some tasks, such as individual-level outreach decisions, individual-level discrete decisions might either be required or be easier to integrate into practitioner analysis pipelines. Improve the continuous model q when possible—e.g. by acquiring more data or changing the modeling approach. This is the predominant approach pursued in the demographic prediction literature (26). However, this approach is insufficient. As we make precise in Theorem 1, improving continuous predictive model accuracy reduces the bias; however, even unbiased Bayes optimal predictors can induce argmax bias, unless predictions have zero error (can perfectly identify the true label for each data point). When discretizing, consider the downstream desiderata—e.g. is accuracy among the labeled set most important, or label representativeness?—and evaluate how label error may affect results. For example, Fraga (7, 9) count voters by imputed race and account for variance induced by the imputation process; they further discuss that bias as we measure here would lead to conservative estimates of differences in voter…
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
As an optimization problem, we formalize this problem of discretizing probabilities and its tradeoffs and find that a joint optimization approach can mitigate these distributional skews with negligible loss in individual prediction accuracy, while using *no additional data*.
Comparison of different discretization methods. Each subfigure shows a 3D probability simplex, where individual points are colored according to the label assigned by the corresponding method. For example, in a), all the blue points (clustered in the bottom left) are assigned the Caucasian label, which has the highest class probability according to the continuous model q for that data point. We use a sample of points from the voter file, and Hispanic, Asian, and Native American probabilities are aggregated into Other (points primarily in the bottom right)
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
e.g., naive discretization methods underestimate downstream disparities: building on work by Argyle and Barber, we find discretization underestimates Black voter turnout in NC, with errors correlating with household income. A continuous estimator is unbiased.

www.cambridge.org/core/journal...
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
These skews can affect downstream tasks, for example discrimination auditing or electoral outreach analyses that rely on either the counts, geographic distribution, or individual labels to be accurate or even unbiased.
Per-county Caucasian bias in North Carolina, under different discretization methods—the underrepresentation of voters of color (i.e. bias overrepresenting the white population) when compared to the aggregate posterior. The most commonly used methods of thresholding and argmax further cause geographic skews—in many parts of the state, very few rows are classified as non-Caucasian, and the bias is largest in counties with an already skewed population. In contrast, Thompson sampling and County-conditional aggregate posterior matching have no bias when compared to the aggregate posterior. Note that geographical correlations in both miscalibration and discretization.
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
We show that the most common, “argmax” discretization method – taking the most likely category – skews the predicted population. In our empirical application, a real-world voter file undercounts the Black voter population by 28% in NC, with further geographic skews.
Bias (undercounting of voters of color) in the voter file. a) In North Carolina where ground truth self-reported data are available, the difference in counts for each group between each discretization method and the ground truth. The argmax method substantially undercounts Black voters in particular, with threshold methods further magnifying such bias. The bar marked “Aggregate Posterior” corresponds to the bias of both Thompson Sampling and Aggregate method; as these methods directly reflect proportions from the continuous model, its undercounting is due to the model’s miscalibration. The additional bias of argmax is thus the bias caused by discretizing the model scores.
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
Now online @pnasnexus.org! Many discrimination auditing and electoral tasks use ML to predict race/ethnicity – by discretizing continuous scores. Can the discretization process cause bias in labels and downstream tasks? Yes! Led by @evandyx.bsky.social

academic.oup.com/pnasnexus/ar...
Paper screenshot. Title: Addressing discretization-induced bias in demographic prediction Abstract: Racial and other demographic imputation is necessary for many applications, especially in auditing disparities and outreach targeting in political campaigns. The canonical approach is to construct continuous predictions—e.g. based on name and geography—and then to often discretize the predictions by selecting the most likely class (argmax), potentially with a minimum threshold (thresholding). We study how this practice produces discretization bias. For example, we show that argmax labeling, as used by a prominent commercial voter file vendor to impute race/ethnicity, results in a substantial under-count of Black voters, e.g. by 28.2% points in North Carolina. This bias can have substantial implications in downstream tasks that use such labels. We then introduce a joint optimization approach—and a tractable data-driven threshold heuristic—that can eliminate this bias, with negligible individual-level accuracy loss. Finally, we theoretically analyze discretization bias, show that calibrated continuous models are insufficient to eliminate it, and that an approach such as ours is necessary. Broadly, we warn researchers and practitioners against discretizing continuous demographic predictions without considering downstream consequences.
February 27, 2025 at 7:43 PM
nkgarg.bsky.social
The entire chat is pretty long/about an ongoing paper we're writing, but here's the relevant portion.

The first paper (Manish's) is the relevant one. arxiv.org/abs/2412.08610. This chat occurred the day after it appeared on arxiv.

(it does get every detail wrong except title, author, id)
January 12, 2025 at 4:56 PM
nkgarg.bsky.social
Also check out Kenny's paper + poster on algorithmic monoculture in matching markets. We find that, with modeling market effects (e.g., applicants also have preferences over firms), intuitions about monoculture being bad for applicants become more nuanced.

Link: arxiv.org/abs/2312.09841
Paper abstract for monoculture in matching markets: Algorithmic monoculture arises when many decision-makers rely on the same algorithm to evaluate applicants. An emerging body of work investigates possible harms of this kind of homogeneity, but has been limited by the challenge of incorporating market effects in which the preferences and behavior of many applicants and decision-makers jointly interact to determine outcomes. Addressing this challenge, we introduce a tractable theoretical model of algorithmic monoculture in a two-sided matching market with many participants. We use the model to analyze outcomes under monoculture (when decision-makers all evaluate applicants using a common algorithm) and under polyculture (when decision-makers evaluate applicants independently). All else equal, monoculture (1) selects less-preferred applicants when noise is well-behaved, (2) matches more applicants to their top choice, though individual applicants may be worse off depending on their value to decision-makers and risk tolerance, and (3) is more robust to disparities in the number of applications submitted.
December 11, 2024 at 4:21 PM
nkgarg.bsky.social
This seems fantastic! "The primary objective was to address current challenges and advance the ongoing discourse on the evaluation of recommender systems. The participants’ diverse backgrounds and perspectives on evaluation significantly contributed to the discourse on this subject."
Screenshot of first page of the report, titled, "Evaluation Perspectives of Recommender Systems: Driving Research and Education"
December 2, 2024 at 5:17 PM