anime-scout

次は何を見る？ · what to watch next

So, what anime should you watch next?

Ever finished an anime and wanted the same feeling again, just with different characters and a different story? Tell the Scout a few anime you've watched and it will find the next one. Or open the AniMap and wander.

Anime–with 1,000+ members on MAL
User ratings used––
Hit rate @10––
Taste groups–clusters on the AniMap

The Anime Scout

Add a few anime you've watched (and liked). The more you add, the better it gets.

Not sure? Try or

Type
Year
Min score

Your list is empty.

Search for an anime above and add it to get recommendations.

The AniMap

Every box is sized by members (popularity) and shaded by MAL score. Click a group to look inside, click an anime to see it.

score 7 → 8
Pick a box to see what's in it.

How it works

The first version of this (2021) used one big similarity matrix from an old Kaggle dump. This one is fresher, mixes in what the anime is actually about, and fits in a few MB.

  1. The data. About 12,000 anime with 1,000+ members from a 2025 MyAnimeList scrape, with detailed tags and a sample of real user ratings from the User Animelist Dataset, and synopses and franchise links from the AniList API.
  2. What it's about. TF-IDF on the synopsis and tags, plus genres, themes, demographic and studio, each with a weight. Two anime are close if they are about similar things.
  3. Who likes it. Like the old Scout, every rating is compared to that user's own average. An anime counts as liked if it's at or above it, and two anime are close if the same people liked them (cosine similarity).
  4. A bit of quality. A small nudge towards anime that are popular and well rated, so the top picks aren't obscure noise.
  5. Your picks. Each anime keeps its 30 nearest neighbours. Your picks' neighbour scores are added up, and the highest totals win.

Does it actually work?

I hid 20% of each test user's liked anime (rated 8+), gave the Scout 3 of the rest, and checked how often one of the hidden ones landed in the top 10. These users were never used to build the model. "New franchises only" ignores sequels of the picks, which are too easy to guess. That column with the last row is what you get with the site's default settings.

MethodHit rate @10New franchises only