— — — —— — —
×
— —— —
— — — —— — —
×
— —— —
— — — —— — —
×
— —— —
← darkroom
AniReco — recommendations from your AniList (feedback welcome)
Hi everyone,
I shipped a small project called AniReco: you enter your AniList username and it suggests anime you haven’t completed yet, based on your list (genres/tags, ratings, and similarity search over a backend index).
Try it: https://anireco.app/ (new domain!)
https://anireco.up.railway.app/ (old domain)
How it works (short version): your list is loaded via AniList in the browser; the heavy scoring runs on my server. Results are cached for a while so repeat lookups aren’t hammering everything.
This is my first thing I’m putting in front of real users, so I’d genuinely appreciate any feedback: weird recs, edge cases (empty list, huge list, private profile if that applies), or “this one line confused me.” There’s a Feedback link on the site if you want to email notes.
Thanks for reading, and hope it’s useful for someone’s backlog :))
— Khushal
Okay whatever algorithm you're using seems to not reduce the recommendation strength for stuff with less liked genres/tags (for example, I am getting a lot of ecchi recommendations)
...Say, do you only fixate on the count of entries with certain genres/tags and not the scores of those entries?
The recommendations for me fixated on male/female protagonist, ensemble cast, comedy and drama as reasons (the only one there that actually matters is ensemble cast) while genres and tags like mystery, psychological and band were nowhere to be seen...
hey @Rinko thanks for the feedback!
The explanation for the recommendation will need to be tweaked but the main recommendation algorithm is building a like and dislike vector according to user's watched anime and what score they've put
I suppose I can add a include / exclude genre filter if that would help get better recommendationsThe way it's calculated definitely needs some work, then... in general the recommendations were really random with next-to-nothing that looked anything like what my preferences clearly are...
Placing less value on exceptionally common genres and tags (drama, the protagonist tags) in the algorithm might help.And indeed a way to filter out results with certain genres/tags would be useful, though ideally shows with genres/tags one is adverse to shouldn't crop up that much in the first place.
Hello @Rinko, thanks again for your precious feedback
I have made some major changes in the recommendation algorithm, weighted tags, multi vector similarity and better balance of signals
Please do try it out, and let me know what you think of it!Thanks !
Ah, based on the numbers of certain tags/genres it seems like you haven't filtered out duplicates from custom lists ^^' I've no clue how muvh that impacts things but probably does to some extent.
Also, shows in planning are taken into account for the recommendation point of "similar feel to"...? To the point most shows name one... I'd expected it to just be decently high scoring watched shows named there... two, no, three here are recommended as the sequels to shows on my planning, basically, and another is recommended for a show on my planning that's related to something I despised.On the whole the tag and genre matches seem much better, and even with these flaws (and the amount of Hibike but fair enough) it has actually helped me a fair bit~
- continue this thread →
Hello @Rinko I am glad anireco was able to help you!
Your feedbacks are super helpful and I really appreciate them :))
I didn't know about the custom list feature, I have added deduplication to fix tags multi counting due to custom lists
Planning shows are not taken into account now, I have added filter and preferences options, where sequels is off by default
In case anyone wants to see sequels anireco will only recommend sequel if user has a completed entry as the last node in a franchise, and not if they have a series in that franchise as planning / watching etcI hope you like the new changes!
Thanks!
Hi! I gave AniReco a go and it definitely has potential. A bit of feedback though:
It's probably worth excluding prequels / sequels / alternatives for shows a user has dropped, or more generally consider the status that the user has given a certain anime. Was surprised to see Mushoku Tensei S2 on there when I dropped S1 really early, for example. I guess this would also cover more positive scenarios where you might factor in number of rewatches or planned status for recs.
Other thing related to the above - don't recommend sequels if the user hasn't watched the original show. It's not a big deal for the shows I got recommended but subsequent entries in a series often spoil the conclusion or key plot points of the season prior in the synopsis. You could also consider excluding sequels for recommendations as a whole, since you probably don't need something to tell you to watch season 2.
Lastly small thing but sometimes the "why we recommend this" info repeats the AniList rating despite that info already being displayed with the other anime info. Screenshot below, but also because it made me laugh a bit:
Hopefully some of that feedback is useful ^^
Hello @adarsh ! Thanks for your feedback!
I have now excluded sequels by default, there might be edge cases but straightforward sequels, alternatives should be excluded by default unless you turn the sequels toggle on
Cleared up the recommendation reasons copy, and also added persistence
Your "Matches your interest in ..." did make me laugh as well :pI hope you find anireco even better after these changes!
Thanks!
Cool idea and I got some stuff that does seem up my lane. Never heard of Wixoss or Kirameki Project! Fun.
(Also, the fact I don't get anything by Fujimoto in the first 10 or 20 recommendations is a HUGE plus - honestly tired of seeing his hacky garbage recommended next to every other seinen I love. Brownie points right off the bat, my man)
My only issue - it seems to take stuff from my Planning list, too. I think it'd be neat if we could switch between "stuff similar to titles on your completed list" and "stuff similar to titles on your planning list" or something. Still, this is neat and has potential, imo.
idk if ppl already pointed these (am not reading all the comments xD) but u should probably not include sequels or prequels to shows already on planned list. Also, consider addding some sort of filtering tools. gl o7
Hello @Nurex the sequel issue should be fixed by large now
I have added some basic filtering by show type for now which should help
Will definitely add more filtering
I hope you like the new changes!
Thanks!nice
am watching a recommendation from your site, I must say it is somewhat accurate, good job
also, when i said filtering, the one I most needed using ur site is avg score, consider it if you haven't already o7
Hii!! I like the idea a lot and it's something I've been missing for a long time. So I hope you continue to work on it and make it reach its potential.
As for feedback:
Your scores for Psychological titles tend to run higher than your usual picks (100 on your list)
From this line I assume the system normalises the user's scores to account for different people having different scoring patterns, is that correct? For instance my scores are quite inflated by design, so while an 8.5 might be a high score for most people, it's below my average. So just to reiterate, do you consider what shows are below the user's average score or something that considers their scoring pattern?
As for format, it'd be nice to have a way to filter which formats the user is interested in. For example I got a lot of movie recommendations when I'd prefer to have more TV recommendations.
About sequels. I did get a lot of recommendations for sequels of series I never watched. I know that filtering out sequels isn't a simple task as it's not so easy as to just check if a series has a prequel, as the "first" entry of a series might still have a prequel that was released later. That said, it'd be nice to have recommendations only feature entry points, aka series you can watch without having another season before. While I can't see a perfect method for that, maybe using release date to compare it to its direct relations (prequels and sequels) could work, or even using popularity, as one could expect the entry point to be the one with highest popularity. Best part is that, while this is computationally expensive to do for the whole catalogue, it need only be done once on server side and cached as a permanent list of valid recommendation targets and it isn't user specific.
It'd also be nice to not have recommendations for entries that are sequels, prequels or alternatives of series the user's already watched. For example my second recommendation was
Legend of the Galactic Heroes Die Neue These
which is a remake of one of my favourite shows
Legend of the Galactic Heroes
So yeah it's pretty likely I'd enjoy it and the recommendation isn't wrong in that sense, but it's also a story I've already watched. I understand some people might want this kinda thing tho, so maybe make it a toggleable setting.Lastly I'm curious about how you determine
Strong thematic match with your favorites
You have a list of tags that relate to themes and compare those or something like that?
Anyway great work so far and I hope it continues to get even better.
Hello @BlindTheFirst Thanks a lot for your detailed feedback! Much appreciated!
So just to reiterate, do you consider what shows are below the user's average score or something that considers their scoring pattern?
Yes the system normalises scores for the user, so it accounts for every user's scoring patterns
I have added filtering for formats
I have fixed the sequel issue, there is a sequels toggle which if you turn on, it will give you sequel for a show you liked and have completed and haven't picked any sequel of it, by default this toggle is off
Lastly I'm curious about how you determine
This is a mix of taste vectors, which are built by going through your anime list and seeing what scores you have given to separate anime, we build like dislike vectors and other things, and run filters on anime which match your likes and are far (vector distance) from your dislikes. There are multiple steps but this is a brief idea of it
Thank you so much again! I hope you like the changes!
this is a fun concept! main problem, in my case, is that it keeps recommending me second seasons of shows that are on my planning list which doesn't really make sense. like other users have said, sequels, prequels, OVAs... should definitely be turned off because it is redudant to tell someone they will like an anime based on them liking its first (or any other) season.
I'd be here for a manga version of this, btw!
Hello @halfmetalchemist ! I love your username haha
I have fixed the sequel issue, sequels are off by default and if the toggle is turned on, it should show sequels only if you have completed the last node in the series, not if you have it in planning, watching etc
Manga version is a great idea! I can definitely look into this after I have built anireco to a satisfactory state, I will let you know once I build that
Thanks and I hope you like the changes!
I'm not willing to continue Kokoro Connect, but I am willing to give THIS another chance
SANDAHello @YoungAnimeDoctor, happy that you found something interesting!
Just curious, have you received Kokoro connect OVA / Sequel etc in your recommendation and you have dropped or rated low some of its prequel?
Just trying to figure out pattern if any for a bad recommendation so I can tweak it further!
Thanks for using the service! :))
Not the ova. Just the series. Yeah, I dropped it.
Hey I have made some updates to the website
This is a bit of a non straightforward thing to fix, but I have tried to solve for it
Please try it out and tell me if you feel any issues
There are other features added too for much more control over recommendations such as anilist score filters etc!Nice. No issues
Hi @khushal thank you so much for working so hard to make this for the community, it looks very interesting though I do have some questions about the recommendations it gives me.
The following 20 recommendations it gives me are:
1: Classroom of the Elite
2: Ga-Rei-Zero
3: Hori-san to Miyamura-kun
4: The Case Study of Vanitas
5: Sankarea: Undying Love
6: Call of the Night Season 2
7: SANDA
8: Talentless Nana
9: Black God
10: SHOSHIMIN: How to Become Ordinary
11: xxxHoLiC
12: Dead Dead Demon's Dededededestruction Zenshou
13: Sarazanmai
14: Battle Vixens: Ikki Tousen
15: Rascal Does Not Dream of Bunny Girl Senpai
16: number24
17: Sonny Boy
18: BlazBlue: Alter Memory
19: SHADOWS HOUSE
20: Kyou kara Ore wa!!Out of all these recommendations the only anime that really catches my eye as something I'd enjoy is "The Case Study of Vanitas" which I have on my watchlist and "Call of The Night Season 2" is a sequel to an anime I've already watched, everything else is extremely random and I have no idea whether I'd like it or not but I don't have high hopes for any of these really.
I use the smiley face rating system and I'm confused why the app chooses to recommend me anime that are similar to those I have given a neutral smiley, it gives me several recommendations for similar to "From The New World" which I have given a neutral smiley face, why does the app think I want a bunch of anime similar to From The New World? It recommends me "Classroom of the Elite" as the first thing that pops up because it's similar to "Kakegurui" which I have given a neutral smiley face, why does the app not focus on finding anime recommendations similar to anime that I've given a happy smiley face rating to instead? Thank you for your time.
yo nice app you should add to where we can click the picture and get taken to the page of the anime and make it not add not yet released shows
u make your own algo or huggingface?
Hi @Anmol69 it's a custom recommendation engine I built, with many steps, rules, gates and logics running
I would love to know also about your feedback, if you found any interesting recommendations through it !
im not even sure if there is a fix for this but i feel i should point it out anyway
while sequels are filtered out i got reccomendations similar to sequals. two seperate danganronpa three arcs showed up despite me not having watched the original. i get that all this is being a bit wishy washy and i doubt there is a genuine fix for this but also occultic nine and chaos head came up when i have steins gate on my planning but havent watched it yet (im not really sure how sequal esque they are to be honest), i also got fate zero despite fate stay night also being unwatched and on my planning list
also did get a reccomendation because it was "similar to joker game" even though i have joker game unrated and on hold after 3 episodes
@guisippi understood
The last point is definitely something I will look into as soon as possible
For the rest of the feedback, some things I can try to fix as long as there is some sort of edge between related anime, through PREQUEL, PARENT, ALTERNATIVE etc
For going one step further I think I will need hyper specialised profile building which learns user's taste constantly in the background. This is something I have planned
I will take this feedback into account and try to improve on it as much as possibleThank you!
i respect the effort but as it is rn it seems pretty useless, the recommendations i get must seem good to whatever algorithm you're using but almost none of them are actually appealing to me
Hello @alicepurple !
Thanks for trying anireco out!
Could you please give examples as to what you got recommended and why you didn't like the recommendation
It would help in making anireco even better and more helpful for all users!Thanks!
Is there a way to sort by match percent? I would love to be able to see what the best match for me is.
Hello @ZadjeZDog I have added sorting options to anireco
Alongwith some filters such as average score to be used while generating recommendations
I hope you like them!
Hard to say how useful this is when you don't publish how it's calculated.
As for a potential future feature, having a way to filter out titles with certain tags/genres would help in fixing a lot of the shortcomings that are bound to exist with any recommendation algorithm.