Cx students and faculty had a great showing at the 2018 ACM Conference on Recommender Systems (RecSys) this year. The conference, held on October 2-7 in Vancouver BC, is the premier international forum for presentation of research and techniques in the broad field of recommendation systems.
First, Oath PhD fellows Longqi Yang and Hongyi Wen, along with Cornell Tech collaborator Eugene Bagdasaryan, gave a tutorial on the OpenRec platform. The tutorial, one of six designed to demonstrate the state of the art in recommendation technology to researchers and practitioners, showed how the OpenRec tool can modularize recommendation algorithms for easier development and testing.
Another contribution by Wang, along with Cx faculty Serge Belongie and Deborah Estrin, explores how recommenders using implicit feedback are built using missing-not-at-random ground truth data, and therefore are biased toward recommending trendy items. The paper investigates the level of bias existing in these ground truth datasets, and proposes an unbiased estimator to reduce bias.
A second paper, by Wen, Yang, Estrin, and Cornell Tech postdoc Michael Sobolev, explores how to create recommendation systems that account for people’s privacy concerns. Many users choose to block data collection, or remove previously collected information about themselves. This makes effective recommendations difficult. However, the authors find that overall performance of state-of-the-art collaborative filtering algorithms is not significantly impacted by removing historical data. The findings suggest a potential win-win solution both for services and users.
Finally, another paper investigates how people search and find new content using audio interfaces, such as smart speakers. In an initial study on voice search for podcasts, Yang, Estrin, and other Cornell Tech researchers found that people using voice interfaces explored less deeply into the available podcast options, and were exposed to less new content overall. These findings have important implications for search and recommendation systems beyond the podcast space, such as increasing diversity of high-ranked content and providing new mechanisms to explore long-tail recommendations.