Recommendation as Ranking

A recommender never shows a number to the user; it shows an order. Once that is taken seriously the learning problem changes shape — the model does not need accurate scores, it needs the right items at the top. That is recommendation as ranking.

What the reframing changes
The target stops being an absolute value and becomes a relative judgement: this item should outrank that one for this user. Training data becomes preference pairs rather than labelled cells, which suits implicit feedback, where a click and a non-click are all that exist. The objective follows — pointwise regression gives way to pairwise losses such as BPR and RankNet, or listwise objectives that score a whole ordering — and evaluation moves to rank-aware measures such as NDCG and MRR.

Why it fits this setting

• Scores need only be ordered correctly, so no capacity is spent calibrating magnitudes nobody ever sees.
• Attention decays sharply with position, so an objective concentrated on the first few slots matches what the product is actually paid for.
• Unobserved items can be treated as weak negatives against observed positives, which turns a one-class problem into a trainable one.

Where it sits
Ranking is the second half of the modern two-stage pipeline: retrieval finds a few hundred plausible items, the ranking formulation orders them. Framing the same task as rating prediction is the classic mismatch — respectable RMSE with a mediocre top ten — because half a star of error on an item the user will never reach costs the metric as much as an error at position one.