Re-designing Recommendation on VolunteerMatch: Theory and Practice

          @misc{ scivideos_22916,
            doi = {},
            url = {https://old.simons.berkeley.edu/node/22916},
            author = {},
            keywords = {},
            language = {en},
            title = {Re-designing Recommendation on VolunteerMatch: Theory and Practice},
            publisher = {The Simons Institute for the Theory of Computing},
            year = {2022},
            month = {nov},
            note = {22916 see, \url{https://scivideos.org/simons-institute/22916}}
          }
          

Abstract

In this talk, I describe our collaboration with VolunteerMatch (VM), the largest nationwide platform that connects nonprofits with volunteers. Through our work with VM, we have identified a key feature shared by many matching platforms (including Etsy, DonorsChoose, and VM): the supply side (e.g., nonprofits on the VM platform) not only relies on the platform’s internal recommendation algorithm to draw traffic but also utilizes other channels, such as social media, to attract external visitors. Such visitors arrive via direct links to their intended options, thus bypassing the platform’s recommendation algorithm. For example, of the 1.3 million monthly visitors to the VM platform, approximately 30% are external traffic directed to VM as a result of off-platform outreach activities, such as when nonprofits publicize volunteering opportunities on LinkedIn or Facebook. This motivated us to introduce the problem of online matching with multi-channel traffic, a variant of a canonical online matching problem. Taking a competitive analysis approach, we first demonstrate the shortcomings of a commonly-used algorithm that is optimal in the absence of external traffic. Then, we propose a new algorithm that achieves a near-optimal competitive ratio in certain regimes. Beyond theoretical guarantees, we demonstrate our algorithm’s practical effectiveness in simulations based on VM data. Time permitting, I will also report on implementing an improved recommendation algorithm on the VM platform and present data from our ensuing experimentation. (Joint work with Scott Rodilitz, Daniela Saban, and Akshaya Suresh)