A bit late to the party, but @pfrazee, you might want to check out our latest paper (dl.acm.org/doi/full/10.... (RecSys '25, 🏆 Best Paper). We introduce a simple post-hoc way to mitigate unwanted recs (provably) from user feedback (e.g., "Show Less"), making sure the model respects negative signals!

A big shout-out to my collaborators! @erasmopurif.bsky.social, Emilia Gomez, Bruno Lepri, @andreapasserini.bsky.social and @cristiancantoro.bsky.social.

However, instead of just removing unwanted items (which we show can hurt engagement), we provide a data-driven property to pick and show previously seen items, preserving both safety and relevance!

We exploit conformal risk control techniques to find a threshold to decide which items we need to remove from the user feed. The calibration procedure requires simple binary user feedback (e.g., "Like"/"Dislike"), thus linking directly the user preferences to RecSys behaviour.

We propose a model-agnostic, distribution-free method that uses limited binary human feedback to limit exposure to unwanted content in recommender systems, provably, and not just by filtering, but by replacing "risky" items with items the user has already liked!