Recommenders / Foundations
verifiedImplicit Feedback
Nobody rates things, but everybody clicks. Implicit feedback is what you learn from behaviour rather than from opinions, and its defining difficulty is that you never see a negative: a video you did not watch might be one you would have hated, or one you were never shown.
Treating unobserved as negative is wrong and works anyway, which is why negative sampling strategy matters more than the model. Popularity-biased sampling makes popular items easy negatives and teaches the model popularity; uniform sampling teaches it less and generalises further. The evaluation trap is the same one: sampling a hundred negatives to rank against produces numbers that do not survive ranking the full catalogue.
Optimise a pairwise objective — maximise σ(x̂ sub ui − x̂ sub uj ) for an observed i and a sampled unobserved j — rather than a pointwise loss over a label that does not exist. The estimator's quality depends entirely on the sampling distribution of j, so the negative sampler is part of the objective rather than an implementation detail.
sampled-negatives holds 50% of the budget; rest holds the remaining 50%.
Negatives sampled per positive against the single positive they are ranked against, in items. Drag the negatives up to watch the comparison get harder — and note the catalogue is larger than any number on this bar.
Reviewed by opendroid · 2026-08-18
- arXiv:1205.2618 — BPR: Bayesian Personalized Ranking from Implicit Feedback