Publications
-
Infrequent Resolving Algorithm for Online Linear Programming
Accepted @ Mathematics of Operations Research.
TL;DR Online LP algorithms are either accurate but expensive (re-solving constantly) or cheap but weaker. We bridge the two by re-solving at only a few well-chosen moments: the policy attains constant regret — even in the hard degenerate case — while solving LPs just O(log log T) times, showing the value of resolving at both the start and the end of the selling horizon.
Preliminary version accepted by WINE 2024: Proceedings of the 20th Conference on Web and Internet Economics.
Winner, Best Student Paper Award — Data Science and Decision Intelligence Community of the Operations Research Society of China, 2024.
-
When to Push Ads: Optimal Mobile Ad Campaign Strategy under Markov Customer Dynamics
Forthcoming @ Manufacturing & Service Operations Management.
TL;DR Customers move between an active and an inactive state; ads can reactivate them but also cause promotion fatigue. Modeling this as a continuous-time Markov chain, we show the seller's optimal campaign strategy takes the form of a triple-threshold policy on the time elapsed since the last purchase or ad.
Preliminary version accepted by WINE 2024: Proceedings of the 20th Conference on Web and Internet Economics.
Finalist, Best Student Paper Competition — POMS-HK, 2024.
-
Dealership or Marketplace with Fulfillment Services: A Dynamic Comparison
Manufacturing & Service Operations Management, 26(5): 1860–1877.
TL;DR Should a platform buy and resell (dealership) or host sellers and fulfill for them (marketplace)? Despite very different pricing, inventory risk, fees, and fulfillment timing, the two models converge in revenue once the market is thick — with a one-to-one mapping between their optimal policies. The choice matters most in nascent markets.
-
Optimal Emission Regulation under Market Uncertainty
Naval Research Logistics, 71(8): 1092–1112.
TL;DR Carbon taxes (price instruments) or emission caps (quantity instruments)? Price instruments deliver higher expected social welfare when emission intensity and market uncertainty are both high or both low; otherwise quantity instruments do better. A hybrid helps most for high-emitting industries at intermediate uncertainty.
Preprints
-
Self-Improving Neural-Guided Pruning: A Graph Neural Network Framework for Scalable Mixed Bundle Pricing
Under review.
TL;DR A graph neural network prunes the exponentially large space of candidate bundles down to a small family, and an MILP solver then sets the final bundles and prices. Because exact labels exist only at small scales, the model iteratively trains on its own large-scale solutions — yielding 13–21% higher profit than bundle-size pricing on 100-product instances at roughly 2% of the runtime.
Previous title: Learning to Price Bundles: A GCN Approach for Mixed Bundling
-
Misspecification Loss in Inventory Systems with Correlated Demand
Under review.
TL;DR What does it actually cost to ignore correlation in demand? We quantify this misspecification loss in inventory systems and characterize how it grows with lead time, the correlation coefficient, and the backlog penalty.
-
From Small to Large: A Graph Convolutional Network Approach for Solving Assortment Optimization Problems
Under review.
TL;DR Train a GCN on small assortment instances where the optimum is computable, then transfer the learned patterns to far larger ones. A model trained on 20-product instances recovers over 85% of optimal revenue on problems with up to 2,000 products — in seconds, beating existing heuristics on both accuracy and speed.
-
Asymptotically Optimal Sequential Testing with Heterogeneous LLMs
In progress.
TL;DR When several LLMs differ in cost, latency, and accuracy — and their accuracy is asymmetric across hypotheses — which should you query, and when do you stop? We show the asymptotically optimal policy uses at most two models: mixing between them while the posterior is ambiguous, then switching to a “specialist” as evidence accumulates.