Implicit Feedback Recommendation

Implicit feedback uses events such as clicks, plays, purchases, dwell time, or skips as indirect preference evidence. Unlike explicit feedback, the user did not state a rating; the recommender must decide how behavior maps to preference, confidence, and exposure.

From clicks to preferences

A common transformation is

$$p_{ui}=1\{r_{ui}>0\},c_{ui}=1+\alpha r_{ui},$$

where $p_{ui}$ is binary preference and $c_{ui}$ is confidence. Weighted matrix factorization uses these in a squared loss, while Bayesian personalized ranking uses observed positives to build pairwise training triples.

Worked example

With $\alpha =5$, three plays imply stronger confidence than one play, but a zero still means “unobserved or not exposed,” not “disliked.”

User-item count $r_{ui}$	Preference $p_{ui}$	Confidence $c_{ui}=1+5r_{ui}$	Interpretation
0	0	1	No positive event; keep confidence low.
1	1	6	Some evidence of interest.
3	1	16	Stronger evidence, but still not a five-star rating.
4	1	21	High-confidence positive behavior.

The zero cells are not hard dislikes; they receive low confidence. This distinction is central to sparse utility matrices and to offline replay in online versus offline evaluation.

Caveats

Implicit logs entangle preference with exposure. A click can mean interest, curiosity, accidental tap, or manipulative ranking position. Repeated recommendations create feedback loops, so production systems need exposure logging, freshness controls, and segment-level checks beyond aggregate CTR.

References

• Hu, Koren, and Volinsky, 2008, Collaborative Filtering for Implicit Feedback Datasets
• Rendle, 2012, BPR: Bayesian Personalized Ranking from Implicit Feedback

Nav

Section — Recommendation Systems and Personalization

← Weighted Matrix Factorization Bayesian Personalized Ranking →