Notes
🔍

Pointwise Ranking

The simplest way to turn recommendation into a problem you already know how to solve: treat each user-item pair on its own, predict a number for it, and sort by that number. Pointwise ranking does exactly that, which is why it is the first thing most systems ship and still what most production rankers are.

What the model learns
One example is one pair with one label — a rating to regress, or a click to classify. The loss is ordinary squared error or log loss, and the ranking is a by-product of sorting the predictions. Click-through-rate prediction is the canonical case, and the whole family of deep CTR models — DeepFM, DLRM, Wide & Deep — is pointwise. Nothing in the objective mentions any other item.

Why it survives despite that

  • Calibrated outputs — a predicted probability means something by itself, which pairwise and listwise losses do not provide. Ad auctions and blends of click, watch time and conversion need that number.
  • It trains on anything — standard classification tooling, class weighting, straightforward distributed training over logged impressions.
  • Impressions give real negatives — an item shown and not clicked is a genuine label for the ranking stage to learn from.

Where it falls short
The objective is not the metric. A model can cut log loss by sharpening predictions on items no list will ever surface and gain nothing on NDCG — the mismatch BPR and LambdaRank exist to address. It also assumes negatives are real, so in retrieval, where unobserved simply means unobserved, it is the weaker choice.

Pointwise ranking scores each user-item pair independently — regressing a rating or classifying a click — and produces a list by sorting the predictions. It is the basis of click-through-rate models such as DeepFM and DLRM, and its advantage is calibration: the output is a usable probability, which blends of click, dwell time and conversion require. Its weakness is that the loss never compares two items, so it can improve without reordering the list at all.

One way to build a shortlist is to give every candidate a mark out of ten, then sort by the marks. Pointwise ranking is that approach. The software looks at one person and one item at a time, guesses how likely that person is to click it, and lines the items up by the guess. It is simple, and the numbers mean something on their own — which matters when a service has to weigh a quick click against, say, how long someone actually watches. The weakness is that it never asks the more direct question: of these two, which should come first?