Research
Behavioral Signals in Recommendation Systems
Studying how dwell, re-reads, and dissent shape recommender outputs, and how to make these signals inspectable.
Recommendation systems take dozens of signals beyond clicks. We studied which signals correlate with quality and how to expose them to users and editors.
Study setup
- Mixed methods: log analysis across news and learning apps, plus user interviews.
- Signals tracked: dwell time, scroll stability, re-reads, saves, respectful dissent, and cross-faction replies.
- Comparison across topics with high disagreement (civics) vs low disagreement (hobbies).
Findings
- Re-reads and cross-faction replies predicted constructive outcomes better than raw clicks.
- Dwell alone was noisy; weighting by scroll stability and active input improved signal quality.
- Respectful dissent comments reduced echo chambers and led to more saves by opposing groups.
Design implications
- Expose a simple “Why this?” showing which signals contributed to a recommendation.
- Let users suppress certain signals (e.g., doomscroll time) from influencing their feed.
- Give editors diagnostics: which signals are over-weighted for a topic, and why.
Prototype
We built an inspector panel that lists signals with weights, a toggle to exclude signals, and a preview of how recommendations change. Participants reported greater trust and adjusted their preferences more deliberately.
