Advancements in Models and Algorithms for Management Science
| Autor: | Yao, Yuanfan (Evan) |
|---|---|
| Rok vydání: | 2024 |
| Druh dokumentu: | Diplomová práce |
| Popis: | Management science is an interdisciplinary field that leverages a variety of analytical techniques to inform effective decision-making within businesses and organizations. It is a dynamic field that is continuously innovating as data becomes increasingly available and businesses leverage new digital technologies. As a result, there is a constant need to develop models and algorithms to address unique decision-making settings. This thesis is composed of three independent chapters, each of which proposes novel modeling insights and algorithmic solutions for real-world problems. Chapter 2 studies a mathematical model in online resource allocation where a decision-maker must efficiently allocate a scarce resource to patient and impatient customers. This study is motivated by recent advancements in on-demand online platforms (such as Uber and Instacart) where customers who are patient (e.g., can wait a few minutes for a ride) are offered a discounted price. Under this model, we develop a simple resource allocation policy that has provable theoretical guarantees under a competitive ratio analysis and is also easy to use in practice. Our work supports the managerial intuition that offering discounts for patient customers leads to more robust and efficient resource allocation. Chapter 3 addresses the challenge of organizing a large corpus of documents into an expert-defined labeling scheme without manual annotation or labeled training examples. This work is motivated by a collaboration with a major pharmaceutical company to streamline root cause analysis of deviations in the manufacturing process. In investigating a new deviation, quickly finding related historical deviations is crucial, but such deviation reports are not organized in a way to facilitate this task. This chapter proposes an innovative methodology called Document Classification with Reference Information (DCRI), which crucially leverages the existence of reference information, documents which describe the taxonomy of interest but are not labeled examples themselves. Empirical results show that DCRI can produce highly accurate labels with minimal intervention from subject matter experts. Based on these empirical findings, we develop a mathematical model for the underlying data generating process and propose both numerical and theoretical finds that further justify the DCRI approach. Chapter 4 studies a novel way of generating insights from black-box classification models by deriving simple conditions under which the model predicts confidently. Existing work on explaining binary black-box classifiers typically studies when the model predicts 1 or 0 without accounting for the confidence (i.e., probability) of the prediction. Our work argues that explaining when a model makes confident predictions is more useful to a practitioner as such predictions typically correspond to when a model is more accurate and reliable. We define a novel evaluation metric for black-box explainers which emphasize confident predictions and develop a local-search based methodology to find interpretable lists of if-then rules that optimize for this metric. Evaluation on six real-world datasets suggest that such rule-based explanations are effective at capturing highly confident data points. By targeting highly confident predictions of black-box model, our methodology generates rules that are more useful than existing approaches which only explain a classifier's binary predictions. Ph.D. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|