Decision-making with matrix-shaped data
Abstract/Contents
- Abstract
- Research on matrix-shaped data has focused on conducting statistical inferences from a matrix based on a sparse number of its observations in a high-dimensional setting, with important applications for recommendation systems, revenue management, marketing, and healthcare operations. In response to the increasing popularity of online platforms, experimentation and causal inference, this thesis addresses the following new and pressing problems: how to leverage non-uniform data patterns in a matrix to recommend favorable products and estimate treatment effects more accurately, how to adapt low-rank matrix structure in online experimentation to save costs, and how to better optimize directly from the matrix-shaped data without the estimation interlude. The first part of this thesis starts with an algorithm called NU-Recommend for learning user preferences to optimize recommendation systems using non-uniform matrix-shaped data. It then ventures into the causal inference domain and applies the NU-recommend algorithm to recover missing values in the panel data, leading to more accurate "synthetic control" methods to measure the efficacy of treatment as the missing pattern in the panel data is highly non-uniform. Online learning is state-of-the-art practice in many companies, as they have widely adopted online experimentation such as multi-armed bandit (MAB) to seek the best product for their customers among periodically refreshed product catalogs. This thesis then introduces the Low-Rank Bandit algorithm that can reduce the opportunity cost of online experimentation by leveraging low-rank structure in matrix-shaped data. This allows for a transition from conducting a one-time prediction with offline data as performed by NU-Recommend to conducting prediction online as the data is collected in real time. Finally, this thesis goes beyond the traditional "estimate then optimize" regime which all previous three parts have followed, and instead designs a direct method that goes from data to decision in one step. This method paves the road for future work of going directly from matrix-shaped data to decision making by leveraging the properties of the matrix structure.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2022; ©2022 |
Publication date | 2022; 2022 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Chen, Wanning |
---|---|
Degree supervisor | Bayati, Mohsen |
Thesis advisor | Bayati, Mohsen |
Thesis advisor | Spiess, Jann |
Thesis advisor | Xu, Kuang |
Degree committee member | Spiess, Jann |
Degree committee member | Xu, Kuang |
Associated with | Stanford University, Graduate School of Business |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Wanning Chen. |
---|---|
Note | Submitted to the Graduate School of Business. |
Thesis | Thesis Ph.D. Stanford University 2022. |
Location | https://purl.stanford.edu/mf481rm5095 |
Access conditions
- Copyright
- © 2022 by Wanning Chen
Also listed in
Loading usage metrics...