Three essays on data-driven optimization and causal inference for online platform operations

Ye, Zikun

Permalink

https://hdl.handle.net/2142/121416 Copy

Description

Title

Three essays on data-driven optimization and causal inference for online platform operations

Author(s)

Ye, Zikun

Issue Date

2023-06-27

Director of Research (if dissertation) or Advisor (if thesis)

• Chen, Xin
• Zhang, Dennis J.

Doctoral Committee Chair(s)

Chen, Xin

Committee Member(s)

• He, Niao
• Seshadri, Sridhar
• Wang, Qiong

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Data-driven Optimization
• Causal Inference
• Field Experiment
• Online Platform Operations

Abstract

With the emerging trends in technology, marketplaces, and society, this dissertation consists of three essays and investigates how far we can improve online platform operations. Towards this goal, this thesis develops data-driven technologies, including causal inference, field experiment, machine learning and optimization methodologies to evaluate and optimize the strategies in the contexts of digital platforms and marketplaces. First, we investigate the estimation and inference problem in the multiple-treatment setting. We develop a novel framework combining deep learning and double machine learning to estimate the causal effect of any treatment combination and identify the best one when observing only a small subset of treatment combinations. This proposed method exploits Neyman orthogonality and combines interpretable and flexible structural layers in deep learning. We prove theoretically that this framework yields consistent and asymptotically normal estimators. To empirically validate our method, we collaborate with a large-scale video-sharing platform, and implement our framework for three experiments involving three treatments where each combination of treatments is tested. When only observing a subset of treatment combinations, the proposed method significantly outperforms other benchmarks to accurately estimate and infer the average treatment effect of any treatment combination, and to identify the optimal treatment combination. Second, we solve the cold start problem for online advertising platforms. The challenges come from the prediction of click-through rates under the limited data and advertisers’ dissatisfaction with the performance of their ads during the cold start. Based on duality and bandit algorithms, we propose Shadow Bidding with Learning algorithm with a provable sublinear regret to balance the exploration and exploitation. We also implement the algorithm at a leading short-video sharing platform and conduct a novel two-sided randomized field experiment on a large-scale advertising platform to examine the effectiveness of our algorithm. Third, we solve a class of nonconvex stochastic optimization, which is a composition of a convex function and a random function. For example, the truncation random function is pervasive in revenue management and supply chain management problems. We propose the Mirror Stochastic Gradient algorithm to solve the nonconvex stochastic optimization problem online. Under some technical assumptions, the algorithm achieves epsilon-global optimal solution guaranteed sample and gradient complexities. We also formulate the complex air-cargo network revenue management problem under booking limit control, random demand, random capacity, random consumption, and routing flexibility as a special case of our optimization problem. We demonstrate the superior performance of our method with higher revenue and lower computation cost than other state-of-the-art control policies.

Graduation Semester

2023-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/121416

Copyright and License Information

Copyright 2023 Zikun Ye

Owning Collections