Stability and Fairness of Matches by Recommender Systems
Recommender systems are widely used on online platforms for matching buyers with products, thus their sellers. Cutting down the search cost (for both buyers and sellers) in large platforms they help create markets. But if buyers and sellers knew the preferences and constraints present on both sides, would they still participate in such matches? Or, would they prefer to form their alternative arrangements? Would they consider such matches fair? In this work, we explore these questions theoretically and through computational experiments on real world datasets. (SSRN)
with :
Matching Donors to Philanthropic Causes on Crowdfunding Platforms
If there is one area where the users are not willing to incur significant search cost that is in finding causes to donate to. Typically the potential donors choose among the opportunities that they have passively received—often choosing to not donate instead of searching for the projects that they would like to support. Therefore, there is an opportunity to improve the recipients’ welfare and donors’ utility by matching projects the donors would like to support and bringing them to their awareness. In this project we develop approaches to learn donors’ preferences from their activities at a large cause-based crowdfunding platform. We then study if such learnt preferences can be used to successfully match projects to the donors and its implication for the funds raised and the donors’ utility (paper).
with : Yicheng Song and Allen Li
Learn to Diversify for Personalized Recommendation
How do users create bundles of content for consumption? Can we learn this from naturally observed behavior to design a personalized recommender system that considers interaction among products consumed in a single online session or a day? In this project we model bundle formation as utility maximizing decision of consumers who are heterogeneous in not only their preference towards different types of content, but also how they substitute one type of content for another, and the cost they associate with discovering content. We learn these individual specific characteristics from clickstreams collected from a large media site and apply it to generate personalized recommendations. (paper)
with : Yicheng Song and Elie Ofek