Recommender Systems with Users in the Loop
Recommender systems have come to serve as the “homepage” for users to access informational items such videos, music, books, etc. Existing recommender systems are built on implicit feedback, e.g., user click to learn user preference given its availability in large volumes. However, implicit feedback captures preference on a user’s first impressions and intentions, neglecting experienced and shifting user interests. Moreover, as socio-technical systems, recommender systems have been criticized as amplifiers for “filter bubbles” and “echo chambers” due to the feedback loop they create which reinforces users’ beliefs and preferences. Few user control mechanisms have been proposed as part of the system to combat the unwanted user and social outcomes. Existing approaches to modify recommendations to better conform with users’ intentions are either inconvenient or ineffective. In this presentation, I will introduce two pieces of my previous work that attempt to address the above problems. In the first work, we leverage post-click feedback as a generalization of implicit feedback, for example, skips in consuming music and videos, to improve content recommendations. In the second work, we focus on a specific form of user control, that is time-sensitive data filtering initiated by users, and investigate how such user control would affect recommendation performance.
Hongyi Wen is an Information Science PhD student at Cornell Tech working with Prof. Deborah Estrin. His research focus is on user-centric recommendation systems that incorporate effective and sustainable user-control mechanisms. He is interested in developing socially-responsible personalization systems towards diverse, fair and privacy-aware. As a DLI fellow, Hongyi will study how implicit or explicit manipulations on recommendation outcomes can be mitigated by improving algorithmic and user-facing system designs.