Applied ML systems
Implicit feedback and objectives
Most recommender training data is implicit, such as clicks, watches and skips. Those signals are plentiful but noisy, and the objective you choose decides what the system learns.
A recommender needs to learn what each user likes, and the cheapest source of that is behavior. Ratings are scarce and come mostly from people who like rating things. Clicks, watches, skips, dwell time and purchases are everywhere. The catch is that these implicit signals are noisy, and the objective you pick decides what the system learns.
Concept
Explicit feedback is a rating, a thumbs-up or a survey answer. Implicit feedback is everything else a user does. Implicit data is far more plentiful, but a click means "maybe interested," not "loved."
Walking through one session
A user sees five items. They skip item 1 after a second, click item 2 and watch 30 seconds of it, and never reach items 3 to 5.
The training labels are:
- Item 2 is a positive. The user clicked and watched.
- Item 1 is a weak negative. The user saw it and skipped it. That could mean "not for me," or just a bad thumbnail.
- Items 3 to 5 tell you almost nothing. The user may never have seen them.
Most systems treat unseen items as soft negatives, not as "dislikes." Treating every non-click as a firm no would teach the model that the first screen is all that matters.
Choosing what to optimize
Objectives look interchangeable until you ship one.
| Objective | Proxy signal | Failure mode if you optimize it alone |
|---|---|---|
| Engagement | CTR, clicks | Clickbait and misleading thumbnails |
| Consumption | Watch time, completion | Long, low-quality videos and rabbit holes |
| Conversion | Purchases, sign-ups | Over-recommending what already converts |
| Satisfaction | Surveys, likes, "not interested" | Sparse, and hard to optimize online |
| Long-term value | Retention, weekly active days | Slow feedback, so it needs delayed labels |
Teams usually blend several heads, such as click, watch, like, share and report, and tune the mix with online experiments.
Trade-off
A short-horizon engagement metric gives a clear signal and fast iteration. A blend with a satisfaction or retention term protects you once engagement starts to game the product. Early on, pick one primary metric and a few guardrails you refuse to hurt.
Common mistakes
- Treating non-clicks as dislikes. Most unseen items are just unseen.
- Optimizing clicks when watch time is the product. Clicks reward thumbnails. Watch time rewards content people finish.
- Forgetting the logs are shaped by the old system. What users see depends on the previous ranking. See position bias.
Check yourself
- Why is a skipped item a weak negative and not a firm dislike? Answer: the user may have skipped it for reasons unrelated to taste, such as a bad thumbnail or an interruption.
- Why might a click-only objective hurt long-term satisfaction? Answer: it rewards clicks, which can come from misleading content that people do not enjoy.
Key takeaways
- Implicit feedback is abundant and noisy. A click means "maybe," not "love."
- Unseen items are not firm negatives.
- The objective you optimize decides what the system learns.
- Blend objectives and guard against the failures of any single one.
Further learning
Next, see how collaborative filtering turns these interactions into vectors in matrix factorization.
Related
Recommender systems
How feeds and stores pick a short list from millions of items. A cheap retrieval step, a careful ranker and business rules. Start here for the map.
Collaborative filtering and matrix factorization
Collaborative filtering recommends items from patterns of who interacted with what. Matrix factorization learns a short vector for each user and item, so a dot product predicts affinity.
Ranking and learning to rank
Second-stage ranking for search, feeds and ads. Learning to rank, ranking metrics, LambdaMART, CTR models, position bias and calibration. Start here for the map.