Implicit Feedback
Hardly anyone rates anything. Almost everyone clicks, plays, skips, scrolls and buys. Implicit feedback is preference inferred from that behaviour rather than asked for directly, and it is what nearly every deployed recommender actually learns from.
What makes it a different problem
Explicit ratings arrive with a scale and with negatives — one star means dislike. Implicit data has neither. Only positives are observed, and an empty cell mixes two entirely different situations: the user saw the item and passed, or never saw it at all. Nothing in the dataset is labelled dislike.
How models cope
Negatives have to be manufactured, so the loss changes shape:
- Confidence weighting — treat every unobserved cell as a weak zero and every interaction as a positive whose confidence grows with the count, the approach behind implicit ALS.
- Pairwise ranking — require an observed item to outrank an unobserved one drawn at random, which is what BPR optimises.
Why it matters
Implicit signals arrive in volumes ratings never reach and cost the user nothing, which is why libraries such as implicit and LightFM assume them by default and why rating datasets like MovieLens are routinely binarised for benchmarking. The cost is interpretive: magnitude means confidence, not enthusiasm — a track played forty times on repeat is not forty times better than one played once. Evaluation moves with it, from rating error to top-k ranking metrics.
Implicit feedback is preference inferred from behaviour — clicks, plays, purchases, dwell time — rather than stated in a rating. It holds positives only: an unobserved user-item pair mixes disinterest with never having been shown the item, so no genuine negatives exist. Models construct them instead, either weighting every unobserved cell as a weak zero with confidence rising on interaction counts, as implicit ALS does, or ranking observed items above sampled ones, as BPR does.
Nobody fills in rating forms. But everyone leaves a trail — what they clicked, what they played to the end, what they bought, what they scrolled straight past.
Implicit feedback is what recommenders learn from that trail, instead of from opinions people state out loud. It is abundant and it costs the user nothing, and it comes with a catch: it only ever records what someone did, never what they disliked. A film you ignored and a film you were never shown look identical in the data. The system has to guess at the difference, and that guess is one of the central design problems in recommendation.