Problem
VK's recommender platform for video and text trains its ranking model mainly on time spent. Clickbait, whose title and cover mislead about the content, holds attention for a long time and gets a disproportionate share of impressions, helped by popularity bias. Users watch it 60% longer than other content, yet they give positive survey answers 20% less often. Banning it is not an option: platform rules are fixed, and some users like such content.
Example from the article: a user with one positive interaction, and it was with clickbait, gives no evidence about taste, while a user with a long history full of clickbait likes is a good audience for it.
Solution in brief
Keep clickbait, but show it only to users who want it. The first version requires a minimum number of positive interactions with clickbait. The second predicts a like for each user-document pair with gradient boosting and keeps only clickbait that lands in the top-K% of that user's own candidates.
Questions this case answers
- How to limit exposure of content that misleads about its own substance in a recommender without banning it, by showing it only to users with a demonstrated taste for it
- How to filter candidate items by a predicted positive reaction when users differ widely in their base rates and the time-based training signal favors misleading content
- How to limit clickbait exposure in recommendations without banning it
- How to show clickbait only to users who actually want it
Tools and methods
- Positive-interaction threshold — show clickbait only to users with at least positive interactions with it.
- Like-prediction model — gradient boosting on user, item and pair features (ALS, Transformer).
- Positional thresholds — a per-user cut at top-K% of candidates instead of one global score cut.
- User surveys — "would you like to see similar content?" as the check that filtering works.
Google or a one-time email code. No password.