Name
Abstract | ||
|---|---|---|
The online world is rife with scenarios in which a user must select one from a finite set of alternatives: which movie to watch, which song to play, which camera to order, which website to visit. There is a long history of study of these types of questions in economics, machine learning, marketing, and psychology. However, historically the study of choice was limited to relatively modest data scales. Today, we have access to large-scale datasets providing insights into the choices of large populations of users faced with a wide variety of sets of alternatives. From such data, we are beginning to develop more detailed models of how users weigh alternatives and make selections. I'll begin this talk by covering basic models for choice, along with some standard assumptions made by these models. I'll then cover some more recent work on understanding choice in modern datasets. First, I'll discuss choice in a geographic setting, in which users make selections of restaurants. Features in this setting provide visibility into the actual time cost to travel to one alternative versus another. In addition to this real cost, we may also tease out the implications on likelihood as the set of alternatives at a particular distance becomes larger or smaller based on the local density of restaurants. The marginals of these factors provide poor quality compared to joint models suggested by choice theory. From choice of individual restaurants, I'll then move to a setting in which users consume items repeatedly: repeated purchases of a particular brand of product; repeated listens to a song; repeated visits to a favorite coffee shop; and so forth. I'll describe a simple but accurate model for this problem whose behavior moves between two regimes based on parameter choices: in one regime, users settle on particular favorites and maintain them over time; in another regime, even the most favored item will eventually be abandoned after finite reconsumptions. Finally, I'll move to a more complex scenario of sequential consumption of a range of items, and will show how the theory of discrete choice can be incorporated into the theory of markov processes, requiring a new algorithmic approach to learning the optimal solution. In all of these instances, the learned models include a per-item quality score that may be viewed as proportional to the residual likelihood of selecting the item after other factors in the choice slate have been accounted for. The work described in this talk is partly due to other researchers, and partly joint with various colleagues including Ashton Anderson, Ravi Kumar, Mohammad Mahdian, Bo Pang, Sergei Vassilvitskii and Erik Vee. |
Year | DOI | Venue |
|---|---|---|
2015 | 10.1145/2806416.2806419 | WebSci |
Keywords | Field | DocType |
Discrete Choice, Repeat Consumption, IIA, Geographic Choice | Scale analysis (statistics),Data mining,Rational choice theory,Quality Score,Visibility,Markov process,Finite set,Computer science,Discrete choice | Conference |
ISBN | Citations | PageRank |
978-1-4503-3794-6 | 1 | 0.38 |
References | Authors | |
0 | 1 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Andrew Tomkins | 1 | 9388 | 1401.23 |