Novelty
A recommender can hit every accuracy target by telling you about things you already know. Suggesting the biggest film of the year to someone who has not rated it is a safe bet and almost no help. Novelty is the metric that notices.
What is being measured
Novelty scores how unfamiliar the recommended items are, as a function of how rarely they appear in the data at all. The standard formulation is self-information: an item's novelty is the negative log of the fraction of users who interacted with it, averaged over the list, so a long-tail title contributes far more than a blockbuster. Variants weight by rank and by estimated relevance, so an unknown item buried at position twenty counts for less than one at the top.
What it is not
- Diversity is how unlike each other the items in a list are — three blockbusters from three genres make a diverse list that is not novel.
- Serendipity adds relevance and surprise: novel items the user actually enjoys. Novelty alone is satisfied by obscure junk.
- Coverage is a catalogue-level property across all users, not a per-list one.
Why it is tracked
Accuracy-optimised models inherit popularity bias from implicit feedback and feed it back into the logs they are retrained on, narrowing what anyone is ever shown. Novelty measures that narrowing, and it trades directly against precision — which is why it is reported alongside accuracy and tuned to a target rather than maximised.
Novelty measures how unfamiliar the items in a recommendation list are to the user, standardly computed as mean self-information — the negative log of each item's popularity — so long-tail items score high and blockbusters near zero. It exists because accuracy metrics reward recommending what is already known, hiding popularity bias and the feedback loop it creates. Unlike serendipity, it makes no claim that the unfamiliar items are relevant.
If a friend only ever recommends the film everyone is already talking about, the advice is not wrong — it is just useless. You would have found it anyway.
Novelty is a way of scoring how much a recommendation actually tells you something new. A suggestion for an obscure album scores high; the current number one scores almost nothing. Systems measure it because the easiest way to look accurate is to keep pushing popular things, which quietly shrinks what anyone ever sees. Novelty says nothing about whether you will like the unfamiliar thing — only that you had not met it before.