AI-Powered Anime Recommendations + Stats

At long last, I've collected fresh data and trained a new model. This model contains new anime released up through July 2024, so newer shows should now be showing up in the recommendations.

I've launched the site in preview currently; you can try it out here: https://anime-preview.ameo.dev/

The reason I haven't swapped it over to the main site yet is because some users are seeing rather significantly different recommendations using the new model than they used to. I'm not sure what the cause of this is or if the difference is overall better or worse.

I'd greatly appreciate some feedback from people as to if they feel that the recommendations from the live site (https://anime.ameo.dev/) or the preview site (https://anime-preview.ameo.dev) are better for them.

Thanks for using the site - and I hope the new model gives you some good shows to check out :)

@ameo___ The preview site says it uses the Jan 2023 model with the latest shows from around that time. Can you please take a look?



Thanks for your hard work! How easy can the site take on new titles as additional seasons air?

@mo_lave my bad - I made a configuration mistake last night and the -preview subdomain was pointed at the old site.

It's fixed now - tyvm for letting me know!

@DesuMaiden I know you're probably a troll, but the "AI" part of this really isn't anything crazy or special. It's a simple neural network that learns relationships between anime based on user watch patterns.

The whole site and model is open source as well: https://github.com/ameobea/sprout

It's just a much simpler version of recommendation algorithms that power things like TikTok, Twitter, etc.

At long last, I've collected fresh data and trained a new model. This model contains new anime released up through July 2024, so newer shows should now be showing up in the recommendations.

I've launched the site in preview currently; you can try it out here: https://anime-preview.ameo.dev/

The reason I haven't swapped it over to the main site yet is because some users are seeing rather significantly different recommendations using the new model than they used to. I'm not sure what the cause of this is or if the difference is overall better or worse.

I'd greatly appreciate some feedback from people as to if they feel that the recommendations from the live site (https://anime.ameo.dev/) or the preview site (https://anime-preview.ameo.dev) are better for them.

Thanks for using the site - and I hope the new model gives you some good shows to check out :)

@ameo___

I've been looking for something like this for awhile! Most recommendation sites I've seen based on MAL profiles tend to just factor in whether or not you watched a series, rather than how you rated it, so I'm glad to finally find something that should, in-theory, recommend me stuff based on my actual tastes!

Without having seen any of the anime on this list, because, well, they wouldn't be recs if I'd already seen them, I'd say the non-preview site's recs feel like they'd be "better" for me, for the most part.

Samurai Champloo (which has a lot of the same creative staff as Cowboy Bebop, which I loved) is on the non-preview site's recs, and even says the reason is because I liked Cowboy Bebop, the Cowboy Bebop movie and Baccano! (which, while unrelated to Bebop is an anime I liked for much of the same reasons I enjoyed Bebop).

The non-preview site also recommends Millionaire Detective because I liked Moriarty the Patriot and, without having seen Millionaire Detective, just feels like that's probably a pretty safe bet.

When I turn on extra seasons, the non-preview site's top recommendation (or, at-least, the one at the top of the list; I'm assuming the recommendations are ranked in some way, but of-course I'm not sure. Even if I looked at the open-source code, I'll openly admit I'd be too math-illiterate to understand most of it. ^_^;) is Free! Season 3. And, yeah, I think it's very safe to assume I'll like that one, since I loved the previous 2 seasons. It does show-up on the non-preview version, too, but much lower down the list.


I will say, though, the fact that the site factors in the one anime I dropped without a rating as "Because you disliked..." could be a double-edged sword. In my case, that anime is one I am never going to watch, so it's not exactly fair to say that I "disliked" it. But I know that most people who drop series without ratings are doing so as a way to say they disliked the series, so that's definitely just a me thing.

Both versions of the site place Corpse Princess quite high on the rec list, so I might check that out soon. I've been looking for something to watch soon!


Out of curiosity, I also tried the site with my IRL best-friend's MAL username, since I think I know her taste pretty well and have a fair guess about whether or not she'd like a given series that I've watched. Both versions of the site only return a few anime I've seen when using her username (those being Death Parade, KADO and Hanako-kun), which isn't a large sample size, but I suppose that isn't surprising since we've watched a lot of the same series, lol.
The preview site recommends Hanako-kun for her, and the non-preview site recommends Death Parade and KADO. I imagine she'd enjoy Hanako-kun and Death Parade, but I *think* she'd probably like Death Parade more, for what that's worth. Honestly, I could see her going either way on KADO. I have been surprised by some of her anime rankings before, though, so maybe she would prefer Hanako over Death Parade? XD

@Kasen-Ibara Thank you very much for the detailed info!!

> I'd say the non-preview site's recs feel like they'd be "better" for me, for the most part.

Yeah... that's the feeling I've had myself and the general vibe I've been getting from other people. I'm not sure why that is - the main difference is that I collected more data for the new version and included a larger number of profiles in it. It could be that some characteristics of the data (average rated anime count, average account age, average age of the people behind the accounts lol) changed which resulted in lower-quality results of the model.

It's for this reason that I've held off on swapping to the updated model so far. I still plan on going back and spending more time to figure out the issue, but I've been working on other projects. That being said, Sprout has been seeing a spike in traffic recently so I might prioritize that work in the next couple of months!

----

Anyway, thanks again for your detailed feedback and info!

@Kasen-Ibara Thank you very much for the detailed info!!

> I'd say the non-preview site's recs feel like they'd be "better" for me, for the most part.

Yeah... that's the feeling I've had myself and the general vibe I've been getting from other people. I'm not sure why that is - the main difference is that I collected more data for the new version and included a larger number of profiles in it. It could be that some characteristics of the data (average rated anime count, average account age, average age of the people behind the accounts lol) changed which resulted in lower-quality results of the model.

It's for this reason that I've held off on swapping to the updated model so far. I still plan on going back and spending more time to figure out the issue, but I've been working on other projects. That being said, Sprout has been seeing a spike in traffic recently so I might prioritize that work in the next couple of months!

----

Anyway, thanks again for your detailed feedback and info!