Why Dislikes Don’t Improve Recommendations
Came across an interesting article from The Verge: dislikes don’t work. This is about YouTube but applies to other services too. So, if one hits ‘dislike’ or ‘not interested’ for a particular video, that video indeed disappears from the recommendation feed. But only that specific video, the number of similar videos will not decrease.
This aligns well with my experience: almost every time we incorporated explicit negative signals into the feature set, they barely had any influence. It’s understandable — they are just too sparse (though it would be interesting to know if this is the case with Tinder). Moreover, it’s not always clear theoretically how they should influence: if a user dislikes one item, does it mean the probability of a positive interaction with similar items decreases? I’m not so sure. Because most of the time, users dislike things in topics they are interested in. Therefore, merely adding dislikes into features seems pointless, unless the negative signal is also included in the target.
I recall a related episode from the early days of personalized radio on Yandex.Music. Back then, our primary model was trained to rank positive interactions against five randomly sampled negatives (sampled by popularity, as far as I remember). At one point, we found that if a user disliked a specific track, tracks from the same artist were seemingly recommended even more. And people started to complain. We checked it out — and indeed, the dislike made the scores of tracks from the same artist only increase. The explanation turned out to be simple. If the model needs to choose a listened track from a ranking group, where the rest of the tracks are random, and one track from this group has a dislike for a track by the same artist, it’s not hard to guess — this track is the one that was listened to.
