Explicit Versus Implicit Feedback
Explicit feedback is a stated judgment: stars, thumbs, survey answers, or direct preferences. Implicit feedback is behavioral evidence: clicks, plays, purchases, dwell time, or skips. Recommenders often have much more implicit data, but explicit data has clearer semantics.
Rating error versus ranking loss
For explicit ratings, a model can minimize observed rating error:
For implicit logs, the common conversion is
which is used by weighted matrix factorization. This separates “preference was observed” from “how confident the event count makes us.”
Worked example
Compare three explicit ratings with three implicit event counts:
| Item | Explicit rating | Implicit count | Preference | Confidence |
|---|---|---|---|---|
| A | 5 | 20 | 1 | 7.09 |
| B | 4 | 2 | 1 | 3.20 |
| C | 1 | 0 | 0 | 1.00 |
The explicit mean rating is , and the one-star rating is a negative signal. The implicit conversion only knows that two items were engaged with and one was not, so it must be evaluated differently in ranking and recommender evaluation.
Caveats
Explicit feedback is sparse and can be biased toward users with strong opinions. Implicit feedback is abundant but confounded by exposure, interface design, and feedback loops. Two asymmetries matter before modelling: a positive implicit observation is not a clean preference (a click can be a mistake or curiosity), and an unobserved interaction is not a clean negative (the item may simply never have been shown). This is why implicit models weight confidence and choose eligibility and negative-sampling rules explicitly rather than treating every missing entry as a dislike or assuming every click equals a four-star rating.
References
- Hu, Koren, and Volinsky, 2008, Collaborative Filtering for Implicit Feedback Datasets
- Koren, Bell, and Volinsky, 2009, Matrix Factorization Techniques for Recommender Systems
Nav