Notes
🔍

Explicit Feedback

When a service asks you to rate a film out of five, or to press thumbs-down, it is asking for a judgement rather than watching your behaviour. Explicit feedback is that deliberately stated preference — the cleanest signal a recommender can get, and the rarest.

What makes it different
An explicit rating carries a graded and signed opinion: four stars means more than two, and one star means actively disliked, which no click can express. That gives a well-posed supervised problem, rating prediction, scored with RMSE or MAE against held-out ratings — the setting the MovieLens datasets and the Netflix Prize were built on, and the reason mean-centred similarity and bias terms exist.

Why it is not enough on its own
Almost nobody rates. Coverage of the user-item matrix is a fraction of a percent, and the ratings that do exist are not a random sample:

  • People rate what they chose to consume, and mostly what they felt strongly about — a selection bias that makes observed ratings optimistic.
  • Rating scales are personal, which is why subtracting a user bias term, or centring on the user mean as adjusted cosine similarity does, comes before any comparison.
  • Production systems therefore train mainly on implicit feedback — views, plays, purchases — abundant but unsigned, and keep ratings as a sparse, high-quality supplement. Treating a missing rating as a zero is the classic error: an unrated item is usually one the user never saw.

Explicit feedback is a preference a user states deliberately — a star rating, a like, a thumbs-down — rather than one inferred from behaviour. It is graded and signed, so it expresses dislike as well as interest, and it defines the rating prediction task evaluated with RMSE. It is also scarce and biased towards items users chose and felt strongly about, so production recommenders rely mainly on implicit feedback and treat ratings as a high-quality supplement.

Explicit feedback is you telling a service what you think: five stars for the film, a thumbs-down on the song, a like on the photo. Nothing is being guessed — you said it out loud.

That makes it the most trustworthy thing a recommendation system can hold. It knows the difference between something you loved and something you disliked, which a bare record of what you watched never shows. The catch is that hardly anyone bothers to rate anything, and the people who do tend to rate what they already chose, or whatever delighted or annoyed them most. So these stated opinions are precious, rare and a little lopsided — which is why most systems quietly lean on what you actually do instead.