Title: Pricing and Causal Inference Under Networks
Date: March 31th, 2025
Time: 10:30AM – 12:00PM
Meeting Link: https://gatech.zoom.us/j/97009427244
Yiming Jiang
Operations Research PhD Student
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology
Committee:
Dr. He Wang (Advisor), H. Milton Stewart School of Industrial and Systems Engineering
Dr. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering
Dr. Ashwin Pananjady, H. Milton Stewart School of Industrial and Systems Engineering
Dr. Jinglong Zhao, Boston University
Dr. Christina Yu, Cornell University
Abstract:
Networks often complicate decision-making and system evaluation due to interactions among interconnected nodes. This thesis addresses two significant problems associated with network effects: dynamic pricing in closed ride-sharing networks (Chapter 2) and estimating the total treatment effect (TTE) under network interference (Chapters 3 and 4).
Chapter 2 proposes an approximate linear programming method to mitigate the curse of dimensionality in dynamic pricing. The developed approach provides a convex, tractable model for generating dynamic policies. Theoretical analysis demonstrates that this model produces tighter revenue upper bounds compared to traditional fluid approximation methods, with numerical experiments confirming its empirical effectiveness.
Chapter 3 introduces a mixed randomization framework combining Bernoulli randomization and cluster-based randomization, extending conventional A/B testing methods. This mixed approach yields unbiased TTE estimates under specified conditions. Two methods are developed to implement the proposed design under both known and unknown interference settings, and asymptotic variance bounds for the estimators are derived.
Chapter 4 investigates the pseudo-inverse estimator in scenarios where the interference network is imperfectly measured. The chapter quantifies the resulting bias, establishes a tighter variance bound relative to existing literature, proposes an enhanced variance estimator demonstrating superior performance, and introduces a novel methodology for detecting interference. The effectiveness of these approaches is validated through large-scale real-world experiments.