Bayesian Personalized Ranking (BPR)

Implicit feedback gives you no negatives — only the handful of things a user actually did and an enormous pile of things they did not get round to. Bayesian Personalized Ranking sidesteps the awkward question of what a non-click means by asking a different one: between these two items, which does this user prefer?

The objective
Training data is a stream of triples: a user, an item they interacted with, and an item sampled from the ones they did not. BPR maximises the log of the sigmoid of the score difference between the two, with an L2 penalty, by stochastic gradient descent. Each update nudges the observed item above its sampled partner and nothing else. The scoring function is pluggable — matrix factorisation gives the familiar BPR-MF, but a k-nearest-neighbour model fits the same frame.

Why pairwise changes the outcome
A pointwise squared-error fit on implicit data has to treat unobserved cells as zeros, so model capacity goes into explaining millions of absences that mostly mean "not seen yet". BPR never claims a non-interaction is a negative; it only claims the observed item should rank higher. The objective it optimises is a smooth surrogate for AUC over each user's item pairs, which is much closer to what a top-N list is judged on.

What to watch
The sampler decides how well this works:

• Uniform negative sampling draws mostly obscure items, which become trivially easy once training starts, so gradients shrink and convergence stalls.
• Popularity-weighted or hard-negative samplers restore signal, at the cost of sharpening popularity bias if pushed too far.
• Sampled items the user genuinely likes but has not reached yet are pushed down as false negatives — the known price of the formulation.

It remains a standard baseline, available in the implicit library, Cornac and RecBole, and still competitive against far larger models on top-N benchmarks.