Designing marketplaces and civic engagement platforms : learning, incentives, and pricing
Abstract/Contents
- Abstract
- Many of our most crucial societal interactions are now mediated by algorithmic systems. We buy goods, find work and hire each other, discuss current events, and make public decisions through online platforms. Non-profit and government actors further use such systems to assign kids to schools, organs to patients, and food to food banks. Principled socio-technical system design requires formalizing an objective and understanding the incentives, behavioral tendencies, and capabilities of participants; in turn, the design influences participant behavior. In practice, design decisions are made jointly through the interplay of experimental and data-driven techniques on one hand, and theoretical modeling and insights on the other. In this dissertation, I describe work designing socio-technical systems in two domains, two-sided marketplaces and civic engagement platforms. I demonstrate how to leverage theory-motivated design and robust empirical analyses together to build better systems, filling in gaps at the interfaces of these approaches. Part I considers the design of surge pricing that is incentive compatible for drivers in ride-hailing platforms. The work compares two potential driver payment policies for such platforms: a new driver surge mechanism (now deployed across the US), in which the surged component of a trip payment is additive (independent of trip length) as opposed to multiplicative (proportional to trip length), the historical standard. The paper presents the theoretical foundation that informed this change at Uber. We model surge evolution as a continuous-time Markov chain; we show that, with multiplicative pricing schemes, strategically rejecting certain trip requests may maximize an individual driver's earnings, to the detriment of others. We then develop an incentive compatible pricing scheme with an approximately affine, closed-form expression. Finally, we analyze counter-factual earnings from more than 500000 ride-hailing trips, validating that our proposal would increase incentive compatibility and earnings stability in practice. Part II tackles rating system inflation on online platforms, studying how to choose the multiple choice question asked of raters. Each potential question induces a joint distribution between the seller's underlying quality and the ratings they receive, and we develop a large deviations based framework to quantify how quickly the true ranking of sellers is recovered, given this joint distribution. With an experiment on a large online labor market, we show that various rating questions yield dramatically different rating distributions: while 80.6% of freelancers receive the best rating using a traditional numeric scale, less than 35.8% receive top ratings using other designs. Our theoretical framework quantifies the resulting information gain and provides a principled way to choose among the scales given the behavioral data. We further show how informational priorities (identifying the absolute best items versus separating unacceptable from acceptable items) should affect design. Part III considers the design and building of systems for participatory budgeting. A key challenge in such systems is to design the elicitation mechanism: participants must be able to share their opinions in a manner that is simple, expressive enough for decisions that lie in high-dimensional spaces, and yet enables provably efficient aggregation. First, I present a new method for people to collectively make a decision on a societal budget. In our method, voters are sequentially asked for their ideal budget within a constraint set determined by the previous voter's answer. This process simulates stochastic gradient descent, and the asymptotic output provably maximizes societal welfare in certain settings. We test our method by building a intuitive user interface and running elections on Amazon Mechanical Turk. Second, we show how to optimize an existing elicitation mechanism -- K Approval, in which each voter identifies their favorite K candidates -- in a principled manner. With real voter data from over thirty elections, we demonstrate that many multi-candidate elections that select W winners are run sub-optimally: whereas voters are typically asked to identify their K=W favorite candidates (e.g., K=1 in a winner-takes-all election), it is learning rate optimal to ask voters to identify their favorite M> K candidates
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 | 2020; ©2020 |
Publication date | 2020; 2020 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Garg, Nikhil | |
---|---|---|
Degree supervisor | Goel, Ashish | |
Thesis advisor | Goel, Ashish | |
Thesis advisor | Johari, Ramesh, 1976- | |
Thesis advisor | Prabhakar, Balaji, 1967- | |
Thesis advisor | Sadigh, Dorsa | |
Degree committee member | Johari, Ramesh, 1976- | |
Degree committee member | Prabhakar, Balaji, 1967- | |
Degree committee member | Sadigh, Dorsa | |
Associated with | Stanford University, Department of Electrical Engineering. |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Nikhil Garg |
---|---|
Note | Submitted to the Department of Electrical Engineering |
Thesis | Thesis Ph.D. Stanford University 2020 |
Location | electronic resource |
Access conditions
- Copyright
- © 2020 by Nikhil Garg
- License
- This work is licensed under a Creative Commons Attribution 3.0 Unported license (CC BY).
Also listed in
Loading usage metrics...