Recommenders and Ranking

It describes each item by its own features, builds a profile vector from the features of items that user liked, and ranks candidates by similarity between profile and item vector. No other user's data is involved.

Explicit ratings are stated preferences on a scale and can express dislike. Implicit signals only record actions, so they are abundant and cheap but noisy, one-class, and count-valued: no interaction is not a stated dislike.

No. A ranked list depends only on the order the scores induce, and any strictly increasing transform of every score leaves that order identical. The wrong numbers cost you elsewhere: thresholds, expected-value arithmetic, anything reading the score as a probability.

Matrix factorization learns a short vector of latent factors for every user and every film, then predicts a rating as the dot product of the two vectors plus bias terms. Every empty cell can be filled that way.

User-based collaborative filtering finds people whose ratings resemble yours and recommends what they liked. Item-based finds items rated similarly to ones you already liked. Same rating matrix, but similarity is computed between rows instead of between columns.

Because both numbers can come entirely from the right-hand item being popular. If milk sits in 70% of baskets, any rule predicting milk reaches high confidence by default. Lift near 1 and leverage near 0 expose it.

An item's latent vector is learned only from that item's observed interactions. With none, the training objective contains no term for it, so the vector carries no information and every cold item scores alike.

Catalog coverage is the share of the catalog that ever gets recommended. Intra-list diversity is how unlike each other the items inside one list are. Novelty is how unfamiliar an item is to the user. Serendipity is relevant plus unexpected.

An unobserved cell mixes two situations that look identical: the item was never shown to the user, and the item was shown and ignored. Only the second is evidence against, so the cell is missing information, not a recorded dislike.

Pointwise fits each item independently against its own label. Pairwise trains on two items from the same list and penalises the wrong order. Listwise optimises a whole ranked list at once. Only the last two target order.

Use alternating least squares when the per-user and per-item solves can be spread across a cluster, or when the loss covers every cell. Use SGD when ratings are sparse and observed-only, and you want a cheap, easily extended update.

Raw cosine treats every star rating as a positive quantity, so two raters who use the scale differently look similar even when they disagree. Subtracting each rater's own mean first turns ratings into signed deviations, so disagreement becomes negative similarity.

RMSE averages prediction error over items the learner already interacted with, weighting them all equally. A top-10 shelf is decided only by which handful of items score highest, so RMSE can fall while the ten shown items get worse.

The examination hypothesis says a logged click happens only when the user both examined that slot and found the item relevant. Since examination falls steeply with rank, a top slot's higher click rate reflects position, not quality.

Adding an item to an itemset can only reduce the number of baskets containing it, so any superset of an infrequent itemset is also infrequent. Apriori works level by level and discards candidates whose subsets already failed.

Give each item feature its own learned vector in the same latent space and build the item's vector by summing the vectors of the features it has. A brand-new item then inherits a position the moment its metadata is known.

The signal splits in two: a binary preference, one if any interaction happened, and a confidence from the count, such as c = 1 + alpha * r. The count scales how much the loss cares, not how much the user likes it.

Bias terms absorb the part of a rating that is not about taste: the overall rating level, how generous the reviewer is, and how well-liked the film is. Factors then model only the leftover interaction.

Divide each logged click by the probability that its slot was examined, so a rank-8 click counts far more than a rank-1 click. In expectation that recovers relevance, but rare deep clicks make the estimate noisy.

Filter on absolute joint basket counts, not percentages; rank by lift with a leverage floor; drop rules that are redundant given a shorter rule; then test the survivors, because a co-occurrence rule does not predict what happens when you intervene.

Combine three levers: infer from context the request already carries such as locale and referrer, elicit taste with a short onboarding prompt, and fall back to a trending list. Decay the elicited prior as behaviour arrives.

The training log only contains items the previous model chose to show. Categories it stopped surfacing collect no engagement, so the next model sees even less evidence for them and shows them less again. The loop compounds every retrain.

Draw a few unobserved items per observed interaction, often around four, and redraw them each pass. The distribution is the real decision: uniform draws are almost all long tail and too easy, popularity-weighted draws are harder but push popular items down.

Form pairs only within one query's list, never across lists, and never enumerate them all: 500 candidates give 124,750 pairs. Keep pairs whose labels differ, sample and weight the rest, and normalise so long lists do not dominate.

A similarity is an estimate from a sample, and three co-rated items is a sample of three. Perfect agreement that small happens by chance, so such pairs flood every neighbour list. Shrink similarities toward zero by co-rating count.

Most recent is per learner, not global, so training still holds events dated after some learners' held-out items. The model absorbs future popularity, co-occurrence and catalog changes it could never have at serving time. A fixed-date cut removes that.

Perturb the order for a small traffic slice so the same items land in different slots, then compare their click rates across slots. Holding the items fixed makes the ratio an examination ratio rather than a quality difference.

Treat an offline top-k win as a screen, not a decision. It scores re-ranking of behaviour that already happened, not behaviour change. Calibrate the bar against your own recorded history of offline versus online deltas rather than a threshold someone invented.

It compresses the transactions into a prefix tree in two passes, then mines recursively from conditional sub-trees built for each item. No candidate itemsets are enumerated and no extra pass over the raw data is needed per itemset length.