Neha Sharma

Haosen Ge, Neha Sharma, Hamsa Bastani, Osbert Bastani (Under Revision), Rethinking Algorithmic Fairness for Human-AI Collaboration.

Abstract: Existing approaches to algorithmic fairness aim to ensure equitable outcomes \emph{if} human decision-makers comply perfectly with algorithmic decisions. However, perfect compliance with the algorithm is rarely a reality or even a desirable outcome in human-AI collaboration. Yet, recent studies have shown that selective compliance with fair algorithms can \textit{amplify} discrimination relative to the prior human policy. As a consequence, ensuring equitable outcomes requires fundamentally different algorithmic design principles that ensure robustness to the decision-maker's (a priori unknown) compliance pattern. We define the notion of \textit{compliance-robustly} fair algorithmic recommendations that are guaranteed to (weakly) improve fairness in decisions, regardless of the human's compliance pattern. We propose a simple optimization strategy to identify the best performance-improving compliance-robustly fair policy.  However, we show that it may be infeasible to design algorithmic recommendations that are simultaneously fair in isolation, compliance-robustly fair, and more accurate than the human policy; thus, if our goal is to improve the equity and accuracy of human-AI collaboration, it may not be desirable to enforce traditional algorithmic fairness constraints. We illustrate the value of our approach on criminal sentencing data before and after the introduction of an algorithmic risk assessment tool in Virginia.

Neha Sharma, Maya Ganesh, Debjit Roy (Working), Partial Information in Fork-Join Operations: Evidence from Multi-Brand Cloud Kitchens.

Abstract: Multi-brand cloud kitchens offer variety by co-locating multiple restaurant brands, allowing customers to order from multiple brands in a single order. To simplify ordering, platforms also offer pre-designed dish combinations, created either as platform bundles or by brands listing one brand's dish in another brand's menu. Our dataset of 6.24 million orders from a multi-brand operator in Asia (68 kitchens) suggests a significant operational penalty for multi-brand orders, as they are 48\% more likely to be late than single-brand orders. Our field visits and interviews suggest that order fulfillment in a multi-brand cloud kitchens is a fork–join queuing process with information asymmetries. Estimates from a reduced-form regression indicate that orders requiring central synchronization across brands, but with information asymmetry for the chefs, have 63\% higher odds of being late compared to single-brand orders. While congestion at cooking stations amplifies coordination costs for all orders, these orders are disproportionately affected. Furthermore, given that the platform hosts both its own and external brands in its kitchens, we find that external brands have roughly half the odds of being late due to multiple contributing factors: simpler menus, better process, and lower congestion. The latter is potentially due to the platform steering demand toward in-house brands by providing them with greater visibility. Our counterfactual analysis reveals that information sharing and prioritization are complements, as prioritizing multi-brand orders enhances performance only when chefs have complete information. Under partial information, FIFO may yield lower delays and better quality outcomes.

Vikas Deep, Neha Sharma, Leann Thayaparan (Under Revision), When Collateral Is a Livelihood: Sustaining Access to Productive Assets.

Abstract: Asset-backed lending unlocks vital credit for underserved borrowers by using the financed asset as sole collateral. Yet, scaling this model remains exceptionally difficult. Lenders must balance low loan values and rapidly depreciating assets against volatile repayment behavior, making it hard to distinguish temporary distress from true default. Furthermore, enforcement actions to redeploy these assets can carry substantial economic and social costs. We address these scaling frictions using data from a large Indian commercial electric vehicle lender serving self-employed workers and small businesses. We formulate the lender decisions as a budgeted restless bandit problem and develop a tractable dual-based priority policy that ranks accounts by the value of immediate enforcement versus continued engagement. We prove global convergence to a unique steady state under a simple verifiable condition and establish asymptotic optimality as the portfolio scales. We also show that the resulting dual-priority policy is interpretable, computationally simple, and exhibits monotone structure under realistic conditions. In observational data, our policy improves on the firm’s current strategy by 1.5% even under highly conservative evaluation assumptions. Empirically, our dual-priority policy outperforms industry benchmarks by more than 15%, and has the lowest false-positive rates as it rarely enforces against borrowers who would have eventually repaid. These gains are driven by earlier enforcement when recovery is unlikely and by greater flexibility for first-time asset buyers and individual microbusinesses.

Neha Sharma and Simin Li (Working), Generative AI shifts technical knowledge production toward recombinant novelty.

Abstract: Public debate about generative AI often centers on the role of humans in knowledge work. We study this question by examining the questions people continue to ask on the largest technical Q&A community, even after the broad availability of LLMs. We find that while overall question volume declines, the questions that persist are those that combine knowledge domains in new ways, referred to as novel questions. More importantly, these novel questions are primarily driven by users combining existing niche domains, rather than new technical domains. This shift in demand for knowledge is also reflected in the hollowing out of the core of the community's knowledge network. Lastly, these shifts are driven by both selection among new entrants and by incumbent users adapting what they ask. Finally, our results raise concerns for the sustainability of the open knowledge and data pipeline that supports the continued development of LLMs.

Description: Will LLMs Replace Coders? Not Entirely.

Neha Sharma, Sumanta Singha, Milind Sohoni, Achal Bassamboo (Under Revision), The Empty Promise: When Do P2P Platforms Fail to Serve Advance-Booking Customers?.

Abstract: Peer-to-peer (P2P) reservation platforms allow customers to make reservations ahead of time. However, the assets are provided by independent \emph{hosts} who decide whether and when to list their assets on the platform, after considering the advance-booking and just-in-time markets. A high advance-booking price can incentivize hosts to list early, but may discourage customers from booking in advance, whereas a low price can fail to attract hosts to list early. We study whether a platform can sustain an advance-booking market and identify the conditions under which it instead collapses to a pure just-in-time market. We then examine whether granting hosts control over both prices and listing times mitigates this collapse. Motivated by data from a large car-sharing platform, we develop a two-period game-theoretic model of a P2P platform that sets the prices for both advance-booking and just-in-time customers. Hosts have private costs and face uncertainty about future asset availability when deciding when to list their assets. We show that hosts list early only if the advance price compensates them for their private fulfillment cost, the expected cancellation penalty, and the continuation value of waiting for the just-in-time market. As the just-in-time market becomes more attractive, there is no advance price that can simultaneously attract early supply and advance-booking customers, i.e., the platform collapses to a just-in-time market. We show that if hosts face a positive cancellation penalty or advance-booking customers have a lower maximum willingness to pay than just-in-time customers, then the collapse occurs at a finite level of just-in-time demand. Finally, we show that granting hosts pricing autonomy does not mitigate the collapse and, in fact, weakly expands the collapse region under mild conditions.

Description:   Media Coverage Kellogg Insight,  ISB Research Bytes. Finalist, Service Science Best Cluster Paper, 2022. Second Prize, CMU YinZOR Best Student Paper Presentation, 2022.

Neha Sharma, Gad Allon, Achal Bassamboo (Under Review), Structuring Online Communities.

Abstract: Online Question and Answer communities were started to supplement customer support services. In contrast to conventional customer support, users in online communities can post questions, and other users with more experience or knowledge can answer these questions. Generally, questions answered get rewards and visibility in the community, while the askers gain knowledge if their questions get answered. We study how users decide to join, leave, and participate in these communities. We link the user participation decisions to the underlying network structure of the community. Finally, we explore the levers a community designer can use to balance user participation level and the community’s efficiency in providing answers to users’ questions. We model the community as a multistage stochastic game where all the users have different skill levels. We find the stationary equilibrium of this game and theoretically show that only a core-periphery network structure can emerge in such communities. This network structure has been empirically observed in most online communities. Furthermore, we find that increasing the cost of asking questions in the community improves the proportion of askers that get answers to their questions. This results in higher user satisfaction. However, a higher asking cost lowers the participation level in the community. This trade-off between participation and community efficiency results in non-monotonicity in the number of users in the community with the participation cost. The paper explores the cost of asking a question as a lever that can be used by communities to control the number and knowledge type of users in the community. The communities typically operationalize higher asking costs by either directly penalizing question asking activity or setting up stricter guidelines for questions to be answered. We find that increasing the cost of asking is not always bad for the community. In fact, a higher asking cost improves user satisfaction which can lead to an increase in the number of users in the community despite higher asking cost. We also discuss how the existence of low knowledge users in the community (and not necessarily the high knowledge users) is essential to the survival of such communities.

Description: Media coverage - Kellogg Insights, Ideas for Leaders Finalist, Service Science Student Paper Competition, 2022. Second Prize, CMU YinZOR Best Student Poster, 2022.

Neha Sharma, Sripad Devalkar, Milind Sohoni (2020), Payment for Results: Funding Non-Profit Operations, Production and Operations Management. https://onlinelibrary.wiley.com/doi/abs/10.1111/poms.13336

Abstract: Payment for results (PfR) funding approach, where donors reimburse the non-profit organization (NPO) based on outcomes, is being increasingly adopted in the non-profit sector. However, there is also concern expressed by many voluntary organizations that such a funding approach puts an undue financial burden on small NPOs and could actually be detrimental to social welfare. In this study, we build a theoretical framework to analyze PfR funding mechanisms. We use a sequential game to model the interaction between the donor and the NPO, with the donor as the first mover. This model captures how PfR funding is typically implemented in practice using social impact bonds (SIB), wherein social investors provide the upfront funding needed by the NPO to implement the project. The donor provides funding, based on the actual benefit delivered, at the end of the project and the investors are paid back using these funds. We find that higher targets set by the donor do not necessarily translate to higher expected utility or expected benefit delivered under PfR. When comparing the performance of PfR and traditional funding (TF) mechanisms, we find that the donor typically has a higher expected utility under the PfR mechanism when the probability of a negative outcome shock is either high or low, and is better off using the TF approach otherwise. When the donor’s opportunity cost of funding the project is high, the donor is better off using a PfR mechanism when her belief about the NPO having low efficiency is sufficiently high. Interestingly, we find that for a large range of parameter values there is a mismatch between the approach that gives a higher expected utility to the donor and the approach that maximizes the expected social benefit delivered. Our model and analysis suggest that the optimal funding approach, and the optimal target set under PfR, depend on the NPO’s financing cost from social investors and project outcome uncertainty.