A recommendation system that only shows you things very similar to what you already consumed can be extremely accurate and still make the internet feel tiny.
Accuracy is not the same thing as discovery.
Recommendation researchers therefore talk about properties such as diversity, novelty, serendipity, and catalog coverage alongside relevance. A system can deliberately spend some recommendation space on material that is less certain but potentially useful.
YouTube’s current discovery guidance says its systems look at what videos are often watched together and may identify videos viewers are likely to watch but have not been exposed to yet. See its Search and discovery tips.
That last phrase matters.
A recommendation does not have to be the statistically safest continuation of the user’s existing habits.
Diversity has more than one meaning
A feed can be diverse in topic but not source.
It can show politics, cooking, gaming, and science while all four come from the same handful of giant publishers. It can show many creators who all express roughly the same viewpoint. Or it can provide genuine source diversity while remaining tightly focused on one subject.
So “more diverse recommendations” needs a defined target.
Are we trying to increase unfamiliar creators? Less-popular items? Different viewpoints? Different languages? Different formats? A broader range of topics?
Those goals can conflict with one another.
Exploration costs certainty
The tradeoff is familiar in recommendation research: exploit what the system already knows works, or explore something less certain to learn more.
Too much exploitation creates repetition. Too much exploration produces a feed full of things the user does not want.
Popularity-bias research shows why this matters. A 2020 study on popularity debiasing in collaborative recommendation found that reducing popularity bias could improve qualities such as coverage and diversity with relatively small losses in accuracy in its experiments.
That does not mean every platform should maximize obscurity. Some popular material is popular because it is excellent.
The important point is that recommendation diversity can be an explicit design choice rather than an accidental by-product.
For a user, deliberate exploration also works outside the algorithm: visit a directory, search by a strange phrase, browse subscriptions chronologically, follow a link from a small site, or intentionally choose a source the default feed never surfaces.
Algorithmic Reality becomes less confining when either the system or the user occasionally asks a dangerous question:
What if the next thing is not more of the same?