Since when were there 'Recommendations' in mangadex?

[undeed]

A nice addition that I didn't know many people asked for.
Seems like a passion side project from the devs?

tho, would've been nice if we could also get the other features like library sorting and filtering

 

[undeed]

Mangadex needs a way more extensive tagging system before a recommendation system based on tag similarity is going to work well. (Look at how many different tags MangaUpdates has.)
At the very least, it needs to take your filters into account. I shouldn't be opening a wholesome manga and being flashed with nearly-naked drawings because both manga have "romance" or "full color". I'm surprised they rolled it out like this.

I'm on the record that I'm among the extreme minority who don't trust and don't like these extensive tags on Mangadex.
However, if these tags are then going to be used in the recommendation algorithm, then that's fine.
I can understand that.

But if that were the case, I would like Mangadex to separate tagging system between actual genres and themes with these "also include" tags.
Because I'm on the side who believes genres and themes have specific meanings and should encompass the majority of the story, and not just because those tags appear once or twice in the story.
(I think someone has already proposed this "3rd level tag" in the past)

 

[saitohd]

I hope they add a disable toggle because I don't give a shit about recommendations and I am worried they will only increase the server requests and cause more temp. suspensions.

 

[Basic_human]

I think a better feature would be giving users the ability to recommend similar manga, like how MAL does it, ૮o.o ა rather than it being automated ૮ .ᆺ. ა

SUCH MAGNIFICENT AND GREAT NEWS

 

[BakedBanana]

I think a better feature would be giving users the ability to recommend similar manga, like how MAL does it, ૮o.o ა rather than it being automated ૮ .ᆺ. ა

"Hey since you liked Solo Leveling, why don't you try out Gal Cleaning? They are quite similar."

 

[Manko-sensei]

They are probably still testing it and it was added because almost every post in the manga subforum is "looking for X manga" or "manga recommendations".

 

[Nemo2342]

"Hey since you liked Solo Leveling, why don't you try out Gal Cleaning? They are quite similar."

They both went to shit in a hurry so that checks out.

 

[Manko-sensei]

They both went to shit in a hurry so that checks out.

And both are about cleaning floors, don't forget that.

 

[Tony_Mushah]

Also it's showing Porn recommendations for normal manga?

It was available in the API since the end of last month actually.
The system is kinda weird. It evaluates the recommendations of given title by "vectoring the title tags, calculates the distance between the vector, and only takes the ones that is in a certain threshold". (if you're familiar with embeddings model then you know what it is)
In a nutshell, the recommendations system only includes tags (content rating, demographic, authors, groups are not included).
For API users though, if you want to filter the recommendations, just pass the recommended title ids to /manga endpoint with the filters.

 

[panzerkampfwagyu]

MOAR TAGS!

Like, the only thing you really have to watch out with that is that you don't have multiple tags for the same thing, really, which I have seen happening on a different site.
But more tags the better, allows for exact searching

 

[Hicleric]

I've stumbled onto what seems like an insular block of recommendations --that is, each of the titles within it recommends the others in the exact same order. I don't see this behavior across most other titles except for ones that recommend this particular block. Curious. I guess it would save processing requirements to have prefab sets of recommendations rather than generating them for individual titles.

 

[Angry_Panda]

I've stumbled onto what seems like an insular block of recommendations --that is, each of the titles within it recommends the others in the exact same order. I don't see this behavior across most other titles except for ones that recommend this particular block. Curious. I guess it would save processing requirements to have prefab sets of recommendations rather than generating them for individual titles.

May be due to title ID of some kind always pulling in a particular sort order. Or not. I'm not a dev/coder/computer program writing guy.

 

[Tony_Mushah]

I guess it would save processing requirements to have prefab sets of recommendations rather than generating them for individual titles.

The recommendation is probably stored in the database and recomputed if a staff or a contributor changes the tags of a given title.

 

[Hicleric]

That would be my guess too. I've seen several titles change recommendations over the course of a couple of hours, so it could very well be updating tags or other relevant factors (sample size: whatever has new chapters today). Too bad, I thought for a while I'd found the secret Sho Habby & Friends master list

Seems like some titles are lacking recommendations and I've seen a couple others feeding into insular blocks (not surprising if it's all structured with vector data as you suggested), so perhaps things are being updated as the new system rolls out.

 

[story645]

In a nutshell, the recommendations system only includes tags (content rating, demographic, authors, groups are not included).

Did you all consider including/weight by list overlaps? Or is the data too noisy?
Something like title X appears on N users reading lists, of the recommended titles Y appears on 90% of lists that also features X, etc? (Basically weighing by/using basket analysis). Or too computationally expensive?

Basically, I'm also curious which algorithm you're using and why.