Skip to content
Tutorials Center
My learning

Topics

Lesson 5 of 38 ยท Concepts and Data

Explicit and Implicit Feedback

Learn the difference between ratings and behavioral signals such as clicks, plays and purchases, and how to turn raw events into usable preference scores in Python.

A recommender learns from what people did. But "what people did" comes in two very different flavors, and mixing them up is one of the most common beginner mistakes. Explicit feedback is something a user tells you on purpose: a star rating, a thumbs up, a written review. Implicit feedback is something you infer from behavior: a click, a play, a purchase, time spent, a search, an add-to-cart.

Most real products have far more implicit than explicit data, because almost nobody rates things. This lesson shows the properties of each, the traps of implicit data, and how to convert raw events into scores that models can use.

Comparing the two

Explicit feedback

Star ratings 1 to 5
Likes and dislikes Clear sign
Reviews Rich but rare

Clear meaning, but scarce and biased.

1 / 2
Property Explicit Implicit
Volume Low Very high
Meaning Direct (like or dislike) Indirect (interest, accident, habit)
Negative signal Yes (low rating) Mostly absent
Bias Self-selection: people rate what they loved or hated Position bias: items shown higher get more clicks
Value of a record Preference level Confidence that the user cares

The missing-negative problem

With ratings, a 1 clearly means dislike. With clicks, a missing click can mean "disliked", "never saw it", or "would have loved it but I was busy". So for implicit data a zero is unknown, not negative. Models for implicit feedback usually treat observed interactions as positive with varying confidence, and treat everything else as weakly negative or sample negatives on purpose.

Position and exposure bias

Items at the top of a page collect more clicks simply because they are visible. A click is evidence of interest and exposure. Logging impressions lets you correct for it later.

From events to preference scores

Raw logs are a list of events. A single user may view a product ten times and buy it once. We need one number per user-item pair.

Python Aggregate events into confidence scores

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Two ideas appear here that return throughout the course: event weights (a purchase counts more than a view) and saturation (the tenth view adds less than the first), here via log1p.

Binarizing explicit ratings

Sometimes you want to treat ratings as implicit: "rated 4 or 5" means liked, anything else means unknown or disliked. This is common when the goal is ranking rather than predicting stars.

Python Ratings to likes

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The threshold is a design decision. A "3" on a generous platform can be a mild complaint; on a harsh one it is praise. We revisit rating scales and user bias in Bias Baselines and Rating Prediction.

Choosing signals thoughtfully

  • Match the signal to the goal. Click optimizes curiosity; completed watch or purchase optimizes satisfaction. A model trained on clicks happily promotes clickbait.
  • Use multiple signals. Combine weak (view) and strong (buy) signals with weights, or train separate heads.
  • Treat time spent carefully. A 10-second view of a 10-second clip is complete; a 10-second view of a 2-hour film is a rejection. Normalize by item length.
  • Beware of accidents. Autoplay and accidental taps create fake positives. Filter very short interactions.
  • Keep negatives when you have them. Skips, dislikes, "not interested" and returns are valuable explicit negatives even in implicit systems.
Python Watch-time ratio as a signal

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Here u1 abandoned the movie (4% completion), but u2 and u3 mostly finished it. Using completion rather than raw seconds avoids calling the short clip's 9 seconds a failure and the abandoned movie's 300 seconds a success.

Many algorithms were designed for explicit ratings, and some work only with implicit data. When you pick a library, check which kind of input it expects, and do not feed it the wrong one.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which statement about implicit feedback is true?

    Implicit zeros are unknown, not negative. That is why implicit models work with confidence, not plain labels.
  2. 2. Why apply log scaling to event counts?

    The tenth view is less informative than the first. Log1p compresses large counts.
  3. 3. Which bias arises because items placed higher on a page get more clicks?

    Visibility alone increases clicks, so clicks mix interest with exposure.
  4. 4. Why use completion ratio instead of raw watch seconds?

    Nine seconds is a full watch of a 10-second clip but a rejection of a two-hour movie.

Summary

  • Explicit feedback is told by users (ratings), clear but scarce; implicit feedback is inferred from behavior, plentiful but noisy.
  • In implicit data a missing interaction is unknown, not negative.
  • Combine events with weights and saturate counts (for example log1p) to get one confidence per user-item pair.
  • Ratings can be binarized into likes for ranking tasks; the threshold is a design choice.
  • Position bias, accidents and mismatched signals (clicks versus satisfaction) are major pitfalls.

Next: The User-Item Matrix