Events & Seminars

We consider a price-based revenue management problem with finite reusable resources over a finite time horizon $T$. Customers arrive following a price-dependent Poisson process and each customer requests one unit of $c$ homogeneous reusable resources. If there is an available unit, the customer gets served within a price-dependent exponentially distributed service time; otherwise, the customer waits in a queue until the next available unit. We assume that the firm does not know how the arrival and service rates depend on posted prices, and thus it makes adaptive pricing decisions in each period based only on past observations to maximize the cumulative revenue. Given a discrete price set with cardinality $P$, we propose two online learning algorithms, termed Batch Upper Confidence Bound (BUCB) and Batch Thompson Sampling (BTS), and prove that the cumulative regret upper bound is $\tilde{O}(\sqrt{PT})$, which matches the regret lower bound. In establishing the regret, we bound the transient system performance upon price changes via a novel coupling argument, and also generalize bandits to accommodate sub-exponential rewards. We also extend our approach to a continuous price setting with contextual information and also a network revenue management setting.

Handloom weaving in India employs as many as 3 million rural artisans and India contributes to as much as 40% of the total handmade carpet exports in the world. Nevertheless, the handmade carpet industry in India remains highly decentralized and inefficient. In this work, we present results from a close collaboration with one of the largest handmade carpet manufacturers in India, to improve efficiency and increase productivity of rural artisans. Our collaborator works with as many as 40,000 rural artisans spread across hundreds of villages in northern India. Using detailed rug manufacturing data, we first perform rigorous empirical analysis to shed light on important factors that affect artisans' productivity. Our work highlights that regular supervision can play a critical role in ensuring timely completion and limiting quality defects, eventually improving artisans' income and productivity. Overall, our findings suggest that one additional day between visits to a loom increases weaving times by 3%-24% on average. Then, using this insight, we develop a novel predict-then-optimize framework for optimizing supervisor visits to artisans and perform numerical experiments to show that this approach can considerably improve productivity and on-time delivery of rugs.

This is joint work with Bier Liu, Somya Singhvi (USC) and Xinyu Zhang (NYU).

Online labor platforms provide freelancers the opportunity to work for clients on a project basis. However, like all projects, disputes can occur when the client and the freelancer cannot reach agreement on the assessment of work quality. Disputes on online labor platforms are traditionally mediated by the platform itself, which is often viewed to be unhelpful or biased. Meanwhile, there are emerging platforms that promise to resolve the dispute with a novel tribunal system and relegate dispute resolution to individual platform users through a voting mechanism. We seek to examine the dispute resolution models used by both the traditional online platforms (i.e., the centralized dispute system) and the emerging online platforms (i.e., the decentralized dispute system), and assess whether such emerging platforms do have the advantage over traditional online labor platforms. Our results provide insights on when and how to adopt the decentralized dispute system.
We find that to ensure a fair voting outcome, the tribunal members should be sufficiently diverse. This indicates that the platform can consider selecting tribunal members randomly, rather than selecting a customized tribunal for each dispute case. Furthermore, our results suggest that the decentralized dispute system outperforms the centralized dispute system only when the freelancer’s skill level is sufficiently high. Thus, gig platforms should consider switching to the decentralized dispute system only if they are able to attest to the freelancer’s skill level (e.g., through certification). Lastly, we show that the decentralized dispute system can be more appealing to policy makers because not only does it move gig platforms closer to a true “sharing economy” by relegating more decision-making power to the individual participants, but it can also induce a more socially optimal outcome.

One of the key issues with recommender systems is the filter bubble phenomenon when consumers are presented only with familiar and repeated types of recommendations, and therefore being isolated in the information bubble. To address this problem and explore consumer preferences, I present a novel approach to providing unexpected recommendations that significantly deviate from consumer expectations, and thus pleasantly surprise them. Specifically, I formulate the unexpectedness objective using state-of-the-art deep learning methods, and then incorporate it into the utility function in a personalized manner that captures heterogeneous consumer propensity to seek product variety. In particular, I demonstrate that it is desirable to provide more unexpected recommendations to variety-seekers, and vice versa. By conducting a large-scale online controlled experiment at the video streaming platform of Alibaba, I show that the proposed model significantly increases various business performance metrics used at the company in comparison to their latest production system. The proposed model has been deployed at Alibaba-Youku serving consumers in the short-video streaming applications.

We study a novel strategy deployed by user-generated content (UGC) platforms to address underprovision. This strategy involves the use of an additional low-barrier content curation tool that allows users to post low-effort content distinct from the high-effort content that is standard on the platforms. It remains theoretically ambiguous whether the introduction of this low-barrier content tool complements or cannibalizes the contribution of high-effort content. By leveraging a quasi-experiment on a large Chinese online question-and- answer (Q&A) platform, we identify the causal effects of the low-barrier content tool on incentives for users to contribute high-effort content (i.e., answers to others' questions) in a difference-in-differences framework. We find that the use of this low-barrier tool complements the contributed answers; adopters of the newly introduced low-barrier content tool increase their volume of contributed answers without compromising the

The Locker Alliance Network (LAN) is a recent smart nation initiative introduced in Singapore for parcel pickup by customers, to improve the efficiency of the last mile operation. This government facility is open to all logistic service providers (LSPs) operating in the country. With more parcels being shifted to locker stations, the number of visits to home locations could be drastically reduced, and the length of the delivery trips to homes will decrease. However, in the case of LAN, the carriers have to substitute these home deliveries with visits to the locker stations, on a separate delivery trip. The challenge is to determine the appropriate size of the LAN (number and location of locker stations), since having too many or too few of these stations may increase the total length of delivery trips instead. Furthermore, given the interoperable nature of the system, how should the government design the network of locker stations to serve all LSPs operating in the city?
We develop a network design model to address these questions. For a given delivery profile, say from an LSP, we first develop a model to jointly minimize the length of the two delivery trips (to home locations and to locker stations). We show that this can be formulated as a Prize-Collecting TSP problem, and reformulated as a second-order cone problem (SOCP) under the logit choice model. We develop a heuristic policy with provable approximation guarantee based on its LP relaxation, for this class of network design problem. Our analysis shows also that there is an optimal number of locker stations needed for efficient operations, beyond which the efficiency of the last mile operations will deteriorate. More importantly, we can use the model to design the interoperable network for multiple LSPs, with possibly different delivery volumes, as long as they have similar footprints. We show that the network expands (almost) in a nested fashion in this case, i.e., the optimal networks for LSPs with smaller scale are (almost) contained in the optimal network for the larger LSPs. Therefore, the optimal interoperable network is very close to the optimal network for the largest LSP, and the optimal density of the locker network is dictated by the optimal density of the largest LSP operating in the country. Participation of the largest LSP is therefore crucial in any government-run interoperable system to increase the efficiency of last mile delivery operations.

Recent social justice movements have rekindled interests in supporting minority-owned businesses, and in response, popular online business-review platforms launched features to mark minority-owned businesses. We ask: Do such markers increase the willingness to visit and support the businesses? In a politically-divided environment, the markers may have heterogeneous effects across those who support the cause of helping minority-owned businesses and those who do not. Moreover, the markers may also act as a reminder of biases and prompt consumers to expect different levels of quality purely based on the ethnicity of the business owner. We conducted two online experiments—using restaurant business as the context—and found that, in aggregate, the markers seemed to have the intended effect (i.e., a positive effect on users’ willingness to visit and support the corresponding restaurants). However, the markers had such an effect only on users who supported the cause of helping minority-owned businesses. For those who did not, the markers had little effect for restaurants that matched positive biases based on the ethnicity of ownership, and worse, the markers actually backfired when the restaurants did not match the biases. We discuss the research and practical implications of our findings.

Although conventional wisdom suggests that product recommendations should benefit consumers, there is a lack of evidence on how they help consumers, specifically, whether product recommendations can algorithmically identify products aligned to consumers’ preferences and thus help them find higher-value products. We estimate the benefit of the collaborative filtering (CF) recommendation system by conducting a randomized field experiment on a US apparel retailer’s website. We collect unique data on the affinity scores computed by a CF algorithm to estimate how product recommendations help consumers search for higher-value products that are lower-priced, fit their tastes better, or both. We show that the discovery of lower-priced and better-fit products are the underlying reasons for higher purchase probability (lower likelihood of failed search efforts) of consumers under recommendations. We further find a higher benefit of recommendations in product categories with higher price dispersion and heterogeneity in consumers’ tastes, which provide additional evidence for these underlying reasons. Finally, we find that consumers substitute other search tools on the website with product recommendations when available. Our findings have implications for online retailers, policymakers, regulators, and the design of recommendation systems.

In this talk, we consider infinite horizon Markov decision processes (MDPs) with "fast-slow" structure, meaning that certain parts of the state space move "fast" (and are more influential) while other parts of the state space transition more slowly (and are less influential). Examples of this type of structure arise in a number of practical applications: multi-product inventory control and pricing, machine maintenance, multi-class queueing, and energy demand response. We propose an approximate value iteration algorithm based on the idea of periodically "freezing" the slow states, solving a set of simpler finite-horizon MDPs, and applying value iteration to an auxiliary MDP that transitions on a slower timescale (and smaller discount factor). We present analyses of the regret of policies generated by our approach, along with empirical results that demonstrate its computational benefits.

The rise of the internet has precipitated the birth and downfall of numerous industries, but perhaps no industry has been transformed to the same degree as news production. Journalism has seen the rise of content aggregation, the proliferation of fake news through social media, and the decimation of local reporting capacity, all of which have served to hollow the newspaper industry. In this work, we examine the downstream effects of this decline industry on an outcome of significant theoretical and practical significance: political corruption. As newspapers are an important investigative arm of local communities, it is possible that the closure of community media will embolden corrupt actors who believe they are less likely to be detected following the closure of a local newspaper. To estimate any effect, we employ a difference in difference approach, exploiting the phased closure of major daily newspapers across the country. Results indicate a significant increase in federal corruption charges in federal districts following closure. Further, we observe no evidence that the rise in online newsvendors and the democratization of the press ameliorates this effect. This highlights the important role of the “fourth estate” in inhibiting corruption in governance and the need to conceptualize the punitive societal effects of the internet more expansively.