When tree-based models, such as Random Forest or XGBoost, are used in optimization, their formulations are often intractable and unscalable. We propose a scalable approximation of the optimization formulation that can optimize over ensemble tree models in linear time while still capturing over 90% of optimality on a variety of datasets. MOTEM (Method for Optimizing over Tree Ensemble Models) is an algorithm for optimizing an objective function that is determined by an ensemble tree model.

AI systems are increasingly demonstrating their capacity to make better predictions

than humans. Yet, recent empirical studies suggest that experts may doubt the quality of these systems. We explore the extent to which a decision maker (DM) can properly learn whether a machine produces better recommendations, and analyze a dynamic Bayesian model, where a machine performs repeated decision tasks under a DM’s supervision. We fully characterize the conditions at which learning fails and succeeds.

We study data-driven operations management problems with high-dimensional data. The standard approach involves two separate modeling phases: learning a low dimensional statistical model (dimensionality reduction) from data and then optimizing a decision problem with parameters input from said statistical model. We propose a prescriptive dimensionality reduction approach that better aligns the two phases and delivers superior results over standard methods (e.g., principal component analysis).

We study a problem of sequential learning and choosing from a group of similar alternatives. The unknown payoff of accepting an alternative depends on a set of common features that enable transfer learning across the group. There is also an idiosyncratic value that needs to be learned by sampling over time. The problem is whether to accept the current alternative, continue sampling, or switch to the next one. We model it as a Bayesian dynamic program and analyze structural properties.

We consider the prototypical problem of sequentially assigning jobs to workers at a large centralized matching platform under an infinite worker supply governed by a fixed distribution; this encapsulates the defining characteristic of large market settings such as online labour marketplaces. The goal is to maximize cumulative payoffs from matches. We resolve several foundational questions pertaining to the complexity of this problem setting and provide novel rate-optimal algorithms and analyses.

In many contexts such as technology and management consulting, clients seek the expertise of providers to find solutions for their business problems. We develop a dynamic game-theoretic model where the provider chooses how to present alternative solutions, and the client chooses which solution to try.

Reinforcement learning is a promising strategy for data-driven sequential decision-making. In many real-world applications, it is desirable to optimize objectives that account for risk in the achieved outcomes. We prove the first regret bounds for reinforcement learning algorithms targeting a broad class of risk-sensitive objectives, including the popular conditional value at risk (CVaR) objective. Our analysis relies on novel characterizations of the risk-sensitive objective and the optimal policy.

Missing information is inevitable in real-world data sets. While imputation is well-suited for statistical inference, its relevance for out-of-sample prediction remains unsettled. We analyze widely used data imputation methods and highlight their key deficiencies in making accurate predictions. Alternatively, we propose adaptive linear regression, a new class of models that can be directly trained and evaluated on partially observed data. We validate our findings on real-world data sets.

We study a newsvendor problem in which the decision-maker only observes historical demands. In contrast to the extant literature, we relax the i.i.d. assumption for past demands and assume instead that they are drawn from distributions within a distance r away from the future demand distribution. We establish an exact characterization of the worst-case regret of Sample Average Approximation. When r is small, we present a near-optimal algorithm which robustifies SAA by using less samples.

Decision-makers often face the "many bandits" problem, where one must jointly learn across related but different contextual bandit instances. We study the setting where the unknown parameter in each instance can be decomposed into a global parameter plus a local sparse term. We propose a novel two-stage estimator exploiting this structure efficiently using robust statistics and LASSO. We prove that it improves regret bounds in the context dimension, which is exponential for data-poor instances.

We consider the problem of maximizing charity donations with personalized recommendations and unknown donor preferences. On charity platforms, a donation is observed only when the recommended campaign is selected by the donor, and an eventual donation is made, leading to selection bias issues. We propose the Sample Selection Bandit (SSB) algorithm that uses Heckman's two step estimator with the optimism to resolve the sample selection bias issue.

We explore a new model of bandit experiments where a potentially nonstationary sequence of contexts influences arms' performance. Our main insight is that an algorithm we call deconfounted Thompson sampling strikes a delicate balance between adaptivity and robustness. Its adaptivity leads to optimal efficiency properties in easy stationary instances, but it displays surprising resilience in hard nonstationary ones which cause other adaptive algorithms to fail.

We study the problem of transforming offline algorithms to their online counterparts, focusing on offline combinatorial problems that are amenable to a constant factor approximation using a greedy algorithm that is robust to local errors. We provide a general offline-to-online framework using Blackwell approachability, producing T^1/2 regret under the full information setting and T^2/3 regret in the bandit setting. We apply our framework to operations problems and produce improved regret bounds.

In this work, we discuss Programmable Actor Reinforcement Learning (PARL), a policy iteration method that uses techniques from integer programming and sample average approximation. We numerically benchmark the algorithm in complex supply chain settings where optimal solution is intractable and show its performs comparable to, and sometimes better than, state-of-the-art RL and commonly used inventory management benchmarks.

Motivated by the successes of deep learning, we propose a class of neural network-based discrete choice models, called RUMnets, which is inspired by the random utility maximization (RUM) framework. This model formulates the agents' random utility function using the sample average approximation (SAA) method. We show that RUMnets sharply approximates the class of RUM discrete choice models. We provide analytical and empirical evidence of the predictive power of RUMnets.

A contingent free shipping (CFS) policy offers free shipment of an order only if it satisfies a pre-specified threshold amount. Our study empirically documents a novel determinant of optimal CFS terms: ease-of-return experience. To reflect its impact on the CFS policy’s embedded trade-offs, a manager shall apply the following counterintuitive adjustment; set lenient (resp. stringent) CFS terms when the customer return process is convenient (resp. inconvenient).

We present our empirical findings on how bundle promotions affect consumer purchase and return behavior compared to markdowns, using a large apparel brand’s in-store purchase and return panel data. In this work, we show that bundle promotions increase the incidence and decrease the return probability of each product compared to products sold with markdowns, controlling for price, discount depth, and item characteristics.

Customers use on-line product reviews more frequently. We explore the impact of product reviews on customer’s valuation uncertainty for an experience product and how that in turn affects a monopolist retailer’s pricing and refund decisions. In a duopolistic competition, the overall sentiment of the reviews are influenced by both retailers. We show that the retailers make their returns more lenient in the presence of product reviews in both settings.

To limit food waste, retailers require suppliers to only send products with a remaining shelf life of at least a minimum life on receipt (MLOR). Such agreements may, however, substantially increase waste at suppliers. We analyze two scenarios of (1) coordinating the supply chain (SC) and (2) collaborating on setting the MLOR level. We show coordinating is neither the only nor always the best way to reduce waste. In some cases, just collaboration can be an excellent way to reduce waste in the SC.

A significant amount of fresh produce is wasted in upstream of the food supply chain due to the high cosmetic standard set by retailers. We examine the economic incentives for retailers to adopt such high standards and their impact on food loss. We show how the retailer’s cosmetic standard decision as well as food loss are affected by rejection rate due to high cosmetic standards and consumers’ willing-to-pay for cosmetic-pleasing products.

Grocery retailers target high inventory levels to avoid out-of-stock situations. A side effect thereof is an undesirable customer picking behavior for the freshest or rearmost item. Products with shorter expiration dates remain at the shelf and convert into food waste over time. Prevailing literature related to food waste in retail neglects this impact. Our paper fills this gap by revealing customer picking behavior in retail stores and by connecting this phenomenon to food waste occurrence.

We examine grocery retailer's selling strategies of cosmetically imperfect produce. We consider three strategies: (1) Discarding of the imperfect produce; (2) Differentiating perfect and imperfect produce and selling at different prices; (3) Bunching strategy where perfect and imperfect produce are sold together. We identify when each strategy is optimal under quality uncertainty and varying degrees of consumer valuations and price sensitivity.

Growing consumer demand for sustainable food and transparent supply chains, makes direct-to-consumer sales an attractive alternative for farmers, next to the retail market. We study how a dual-channel structure can improve an agricultural food supply chain’s profitability and sustainable transformation, considering the effect of random crop-yield, customer appreciation of additional producer information, and esthetical requirements of retailers.