Implicit-Feedback Recommendation
Implicit feedback infers preference from behavior like clicks and plays, where non-interaction is ambiguous.
No stars, only actions
Most real recommendation data is implicit: users click, watch, purchase, or skip, but rarely rate. This differs sharply from explicit ratings. A positive action signals interest, but the absence of an action is ambiguous, it may mean dislike, or simply that the user never saw the item. Treating all non-interactions as negative is wrong, yet ignoring them entirely leaves nothing to contrast against.
Confidence-weighted modeling
A standard formulation replaces ratings with a binary preference (one if the user interacted, zero otherwise) and a confidence that scales with the interaction strength, for example c = 1 + alpha * count. Positive interactions get high confidence; unobserved pairs get low but nonzero confidence, so they gently pull predictions toward zero without being treated as certain dislikes. Weighted ALS optimizes this efficiently.
Ranking-first learning
Because the goal is to order items rather than predict a number, many methods optimize ranking directly. Bayesian Personalized Ranking learns from triples (user, observed item, unobserved item) and pushes the observed item to rank above the unobserved one. Negative sampling makes this tractable by drawing a few unobserved items per positive rather than scoring the whole catalog.
Evaluation that matches the task
- Use ranking metrics: recall at k, NDCG, mean average precision
- Split by time so the model predicts genuinely future behavior
- Account for exposure bias: users can only interact with what was shown
- Watch popularity bias, which ranking metrics can reward superficially
Implicit feedback is the dominant regime for streaming, retail, and content platforms, and it connects directly to collaborative filtering and matrix factorization with the confidence and sampling adaptations above.