Food delivery platforms (DoorDash, Uber Eats, Meituan) are racing to expand by signing as many restaurants as possible. Such expansion directly translates into changes in restaurant density on the platform, making it crucial to understand how density affects restaurant outcomes. This paper examines the impact of restaurant density on incumbent restaurants and compares three key externalities induced by increased density: customer expansion and cannibalization, which are well documented in the literature, and delivery pooling (i.e., drivers being closer to restaurants and bundling multiple orders before delivery), which has received much less attention. Using a queueing-theoretical model, we develop testable hypotheses on how these externalities influence delivery speed, sales, revenues, and customer spending. Employing two-way fixed effects and instrumental variable estimation with data from a major Chinese food delivery platform, we find that higher restaurant density improves delivery speed and sales for partnered restaurants but reduces order size; heterogeneous-effects estimation further reveals that fast-food and high-volume restaurants benefit more from increased density, while made-to-order and thermally sensitive cuisines see weaker gains. Combining the empirical evidence with our theoretical predictions, our evidence is most consistent with delivery pooling playing a more pivotal role than customer expansion or cannibalization in shaping density effects. Furthermore, nonlinear least squares estimation of our theoretical model reveals that driver arrival speed at restaurants rises with density, shortening customers' wait times and inducing them to place more—but smaller—orders. However, this effect follows an inverted-U-shaped pattern, indicating diminishing delivery pooling benefits—and even adverse effects—at very high density. Finally, we extend our theoretical framework to develop a simulation model, estimated using data, which allows us to isolate the effect of each externality and identify restaurant-specific optimal densities through counterfactual simulations. Results show substantial heterogeneity in sales-maximizing densities across restaurants, with most operating below their optima. Our findings also suggest that platforms should prioritize expansion in low- to moderate-density regions to maximize delivery efficiency and restaurant sales, while carefully managing congestion and tailoring support based on restaurant type in denser markets.
Contextual bandit algorithms are useful in personalized online decision-making. However, many applications, such as personalized medicine and online advertising, require the utilization of user-specific information for effective learning while maintaining user privacy due to privacy concerns. This motivates the introduction of local differential privacy (LDP), a stringent notion of privacy, to contextual bandits. In this paper, we design LDP algorithms for stochastic generalized linear bandits to achieve the same regret bound as in non-privacy settings. Our main idea is to develop a stochastic gradient-based estimator and an update mechanism to ensure LDP. We then exploit the flexibility of stochastic gradient descent (SGD), whose theoretical guarantee for bandit problems is rarely explored, to address generalized linear bandits. We further conduct experiments using both simulated and real-world datasets to demonstrate the consistently superior performance of our algorithms under LDP constraints, with reasonably small LDP parameters (ε, δ) to ensure strong privacy protection.
How Demand Allocation Shapes Inventory Decisions Online platforms that fulfill orders for independent sellers often bear inventory costs without choosing how much sellers stock. In the paper “Managing Newsvendors via Demand Shifting,” Yilun Chen and Jiaqi Lu study how reallocating customer demand among substitutable products can influence sellers’ inventory decisions. The platform commits to a demand-shifting rule that effectively alters the demand distribution perceived by the sellers. Its optimal rule reduces demand fluctuations around a target inventory level, making that level attractive to sellers and pooling demand risk across them. The authors characterize this rule in closed form. When shifting demand is inexpensive, the platform can achieve the profit available under centralized inventory control, whereas sellers and customers also benefit. However, when shifting is costly and inventory is expensive for the platform, its optimal intervention can reduce customer service levels. The findings show why platforms should assess how demand allocation changes stocking incentives and customer outcomes.
On the Power of Adaptivity in Choice-Based Matching Platforms Online labor, lodging, and carpooling platforms decide not only whom users can choose but also when those options are shown. Simultaneity can crowd attention around a few popular users and leave many good matches unrealized. In the forthcoming Operations Research paper, “Two-Sided Assortment Optimization: Adaptivity Gaps and Approximation Algorithms,” the authors study how much platforms gain by updating assortments after observing earlier choices. They compare four policy classes, ranging from showing all assortments at once to fully adaptive policies that process users one by one. They prove that adaptivity within one side improves performance by at most a factor of [Formula: see text], whereas fully adaptive policies improve on one-sided adaptive policies by at most a factor of two; both bounds are tight. They also develop efficient approximation algorithms, extend their results to constrained settings, and empirically show that their methods perform far better than their worst-case guarantees.
Learning When Demand Is Unknown Airlines launching new routes cannot wait for years of historical data before deciding how many seats to offer at each fare or what prices to charge. They must learn when operating, often from incomplete feedback: selling 30 economy plus seats does not reveal whether demand was 30, 40, or more. In the accepted Operations Research paper “Semibandit Learning for Monotone Stochastic Optimization,” Rohan Ghuge, Zhengjia Zhuo, Arpit Agarwal, and Viswanath Nagarajan develop a general framework for such decisions. They identify a basic property, monotonicity, that allows an optimization method built for known demand distributions to become a learning algorithm for unknown distributions. The framework applies to airline and hotel revenue management, sequential posted pricing, and optimal-stopping problems. Even with partial, censored, or yes-or-no feedback, its cumulative performance gap grows essentially at the square root rate, best possible up to logarithmic factors. Firms can, therefore, improve decisions as data arrive without pausing operations for a separate data-gathering phase.
