Explicit feedback is preference information that a user provides deliberately and knowingly about an item, such as a five-star rating, a thumbs-up or thumbs-down, a written review or a direct answer to a survey question. In recommender systems, it is the classic training signal for predicting what a person will like, because it states both the direction and the strength of a preference.
It contrasts with implicit feedback, which is inferred from behavior such as clicks, watch time or purchases. Explicit feedback is more precise per data point but far scarcer, and that tension shapes how modern recommendation engines use it.
Forms of explicit feedback
Explicit feedback appears in several forms that differ in granularity, effort required and information content. A rating scale captures intensity, a binary signal captures direction, and free text captures reasons that numeric scores cannot express.
The table below compares the most common forms used in production systems and research datasets.
| Form | Typical scale | Information captured | User effort |
|---|---|---|---|
| Star rating | 1 to 5 or 1 to 10 | Direction and intensity | Low to moderate |
| Binary rating | Like or dislike | Direction only | Very low |
| Written review | Free text | Reasons and aspects | High |
| Pairwise preference | A preferred over B | Relative order | Moderate |
| Preference survey | Selected genres or topics | Initial interests | Moderate, once |
| Negative signal | Not interested, hide | Explicit rejection | Very low |
Many platforms have moved from five-star scales to binary thumbs, because users tend to cluster at the extremes and a simpler control collects more responses. The trade-off is lost intensity: a binary signal cannot separate mild approval from enthusiasm.
How recommender models learn from explicit ratings
Recommender models learn from explicit feedback by fitting a function that reconstructs observed ratings in the user-item matrix and then predicts the missing entries. This framing is called rating prediction and was the dominant research task for more than a decade, popularized by the Netflix Prize and datasets such as MovieLens.
In matrix factorization, each user and item receives a latent vector, and the predicted rating is their dot product plus user and item bias terms. Training minimizes squared error only over observed ratings, with regularization to avoid overfitting sparse data. Missing entries are treated as unknown, not as negative, which is the key modeling difference from implicit feedback.
Because users rate on personal scales, preprocessing matters. Mean-centering per user removes the habit of generous or harsh raters, and some systems normalize by each user’s rating variance as well.
| Aspect | Explicit feedback | Implicit feedback |
|---|---|---|
| Signal meaning | Stated like or dislike with strength | Inferred interest from behavior |
| Negative evidence | Low ratings are true negatives | No clear negatives, only absence |
| Missing entries | Ignored during training | Weighted as weak negatives or sampled |
| Typical loss | Squared error on ratings | Ranking loss such as BPR |
| Typical metric | RMSE or MAE | Recall@K or NDCG@K |
| Data volume | Small share of interactions | Nearly every interaction |
Strengths and weaknesses of stated preferences
The main strength of explicit feedback is unambiguous intent. A one-star rating is a real negative, while a click or a short view may mean curiosity, an accidental tap or clickbait success rather than satisfaction.
The main weakness is extreme sparsity. In a typical consumer catalog, only a small fraction of interactions ever receive a rating, so the user-item matrix of explicit signals is usually far more than 99 percent empty. This sparsity worsens the cold start problem and makes models fragile for users with few ratings.
Explicit feedback also carries systematic biases that experienced practitioners correct for:
- Selection bias: people mostly rate items they chose to consume, so ratings are missing not at random and skew positive.
- Extremity bias: users with very strong opinions rate more often, which produces J-shaped rating distributions.
- Anchoring and herding: a visible average rating pulls new ratings toward it.
- Scale drift: the same user may rate more harshly over time as expectations rise.
- Manipulation: fake ratings in a shilling attack can push or bury items, so explicit inputs need fraud screening.
Combining explicit and implicit feedback in practice
Modern recommendation engines rarely rely on explicit feedback alone; they combine sparse explicit signals with dense implicit signals. Implicit data drives candidate generation and ranking scale, while explicit ratings and dislikes act as high-confidence labels and guardrails.
The table below shows typical ways teams integrate explicit feedback into a multi-stage ranking pipeline.
| Use | How it works | Benefit |
|---|---|---|
| Onboarding survey | New users pick topics or rate a few items | Reduces user cold start |
| Hard negative filter | Items marked not interested are suppressed | Respects user control and trust |
| Ranking feature | Ratings and dislikes feed the ranking model | Improves precision of top results |
| Training label weight | Rated interactions receive higher confidence | Separates satisfaction from mere clicks |
| Evaluation signal | Ratings validate implicit-only models | Checks that engagement reflects quality |
| Review mining | Text is parsed into aspects and sentiment | Supports explainable recommendation |
A careful design asks for feedback at low-cost moments, such as after an episode ends or a delivery arrives, and makes negative feedback easy to give. Systems that ignore a dislike quickly lose user trust, even when offline metrics look strong.
Common misconceptions about explicit feedback
A frequent misconception is that a lower RMSE means better recommendations. Accurate rating prediction across all items does not guarantee good top-N lists, because users only see a few items and errors on obscure titles barely matter.
Another misconception is that explicit ratings are objective ground truth. They reflect mood, context and social influence at the moment of rating, so they are better treated as noisy but informative labels. Finally, average ratings are not a ranking on their own: an item with two five-star ratings is not more reliable than one with thousands of ratings averaging 4.6, which is why systems apply Bayesian smoothing to rating averages.
The simple explanation
Explicit feedback is when you tell a service directly what you think. Giving a film four stars, pressing a thumbs-down button or writing a short review are all examples. You are speaking on purpose, so the message is clear.
The service can also watch what you do, such as what you click or finish, but that is guesswork. Explicit feedback is like a friend telling you their opinion, while behavior is like guessing from their face. The direct answer is clearer, but people rarely give it, so services use both.
How it is used in practice
A mid-sized online bookstore with about 200,000 active customers finds that its recommendations, trained only on purchases and page views, keep suggesting titles that customers buy and then return. Only around 3 percent of orders ever receive a star rating, but those ratings show that many of the frequently returned books were rated one or two stars.
The team adds a one-tap rating prompt in the delivery confirmation email, adds a visible not interested button on every recommendation, and feeds both signals into the ranking model as high-weight labels. Low-rated books are demoted for customers with similar purchase histories, and dismissed titles are suppressed for six months.
Over the following quarter, the rating rate rises to about 9 percent of orders, and return rates on recommended books fall noticeably while the click-through rate stays stable, showing that the explicit signals improved satisfaction rather than just engagement.