Can something be done to improve the recommendations?

[Oobaka-san]

It may be a stupid request, but the recommendations on this site are often laughable, to the point that I suspect most people who make recs don't even read what they're putting recommendations on. This is particularly bad in the mystery genre, an area where I've also had to give up on the genre tags as the genre tag includes huge numbers of series with little to no mystery in them, or with just one arc where there's some search to figure out who the villain is, yet they're still called "mystery" series, cluttering the list and making it harder to find stories that actually are centered on mystery (I think I suggested it before, but a "primary genre" or something like that might be useful to deal with this, as well as with romance which you could add to 90% of series as most series have some romantic aspects).

For example, take a mystery series called Shion no Ou. The story's about a teenage girl's efforts to become a major pro at the game of shogi where normally the strongest female-league player is worse than upper-level students aiming to become pro in the main league, all the while torn by the traumatic memories of her parents' murder when she was five. Along with the sport element as she works to improve, much of the story focuses on the struggle to find the killer despite a surprising lack of clear evidence. There are lots of interesting characters, including another girl who's working to gain recognition in an environment where people normally see girls as "cute" at best, a boy who cross dresses in order to compete in the women's league and make easy wins so he can pay his dying mother's hospital bills, a young man who suspects his beloved's death was connected to the murder mentioned above and engages in his own independent investigations, and several pros with their own situations and objectives. What do all the recommendations have in common with the series then? Psychological elements? Murder mystery? Shogi? Nope. Cross-dressing. Even worse is QED, a series that's mostly an episodic series of murder mysteries solved by an athletic JK and a boy genius, which has only one mystery in its recommendations while most of the remainder is sports series for some incomprehensible reason, with one or two basic slice-of-life stories scattered in between.

These recommendations are supposed to lead fans of a work to similar works or works that people who specifically liked that work could seriously expect to enjoy in a similar way, right? As is, this seems to have failed pretty bad, at least in some places. Might not be noticeable for people whose main interests are mainstream stuff like Shounen battle or Shoujo romance, but if you're looking for some other genres it can be like looking for a needle in a haystack, a haystack made entirely by people thinking putting hay in the "needle" pile is helpful.

 

[Hicleric]

The recommendations aren't curated at all, from what I recall of the initial discussion when they first launched. There's some background logic (probably) related to tags that is refreshed every so often but I don't believe we've ever been granted a peek under the hood. Could be cover style and aesthetics for all I know

I would be interested in seeing utilization rate of this feature but I suspect it's not trackable and also quite low. I've occasionally found some interesting titles worth following up but as you point out, more often than not it doesn't produce actionable leads and I often just forget it exists.

 

[undeed]

I don't know how the recommendation list is actually selected,
but ever since the feature was first introduced, I immediately assumed that it's just displayed based on a similarity sorting of the tags listed for each manga.

So, all the recommendation anomalies you mentioned are very easy to explain.
The similarity likely only uses the simplest calculations like Jaccard, Pearson, or cosine, without any complex recommendation system that uses graphs and relationships between manga, let alone relationships with the readers.

For example, with QED you mentioned, the Pearson similarity value with everything that appears in the recommendation tab is >0.5.
Meanwhile, if you calculate its similarity with something like Hyouka or Master Keaton, the value is only 0.4.
Tho, the strange thing is, CMB manga value is actually quite high, so it should have appeared in the recommendation list too. It might just be because of the sorting process. idk.

And since the arrays being compared are quite long and sparse, (78 tags in total, I think?)
adding with the tag management process itself has been quite inaccurate from the start the similarity values end up being somewhat contaminated.
(which is why I've never trusted tags on MD since time immemorial),

It's okay for just a fun feature.
Just think of it like a more "guided" random manga list

It could be somewhat improved if some weights were applied.
I have also previously suggested implementing levels in the tag system, which could then be used as weight values.
But then it means that someone have to do some tinkering around to figure out the right weights for that.
And that's easier said than done.

 

[RogueKitsune]

Can something be done? Yes
Will something be fine? Unlikely (at least in the near term)

The recommended titles you see on a title page are "chosen" purely based on "similar" tags.

The recommended titles on the home page is an actual curated list.

 

[Oobaka-san]

Can something be done? Yes
Will something be fine? Unlikely (at least in the near term)

The recommended titles you see on a title page are "chosen" purely based on "similar" tags.

The recommended titles on the home page is an actual curated list.

I sort of suspected that might be the case, but if so then the way the engine prioritizes tags is simply broken. Mystery seems to have absolutely zero standing in associations, whereas cross-dressing takes priority over everything. And things like demographics, content rating and "content" tags like gore and sexual violence frequently play no part at all, even when dealing with series that have those as major points.

That's part of why I'd also suggested before putting in some kind of prioritizing or rating system for the tags. Sailor Moon should be labeled as Magical Girl before Thriller. QED should be Mystery before Slice of Life. Uwagaki should be Romance before Sci-fi. And Shion no Ou should be Mystery and Sports before Cross-dressing. Other series go the other direction, more Thriller than Magical Girl, more Slice of Life than Mystery, more Sci-fi than Romance. But clearly with everything being treated as the same level in all cases leaves whatever engines are producing the recs throwing darts at a board while blindfolded and drunk.

 

[Bartolumiu]

The current recommendations system is a really basic system that's purely based on the tag combinations for a given title.
However, that doesn't mean we're going to leave it as-is.
There are a couple of improvements and expansions to the system in the roadmap, but there's no ETA yet due to other features and fixes having higher priority.

 

[Hicleric]

whereas cross-dressing takes priority over everything

I mean...have you read any comments recently?

 

[Mangamoz]

The current recommendations system is a really basic system that's purely based on the tag combinations for a given title.
However, that doesn't mean we're going to leave it as-is.
There are a couple of improvements and expansions to the system in the roadmap, but there's no ETA yet due to other features and fixes having higher priority.

Just be careful if you plan to let users make the recommendations, cause I promise people will just post meme recommendations for lols.

"Reads cute wholesome romcom, gets recommended Grapara because it is technically a romcom"

 

[Bartolumiu]

Just be careful if you plan to let users make the recommendations, cause I promise people will just post meme recommendations for lols.

"Reads cute wholesome romcom, gets recommended Grapara because it is technically a romcom"

I don't have the specifics, but I'd assume it'll work in a similar way to how recommendations work in other platforms, where using the preferences from all the users, a map is generated where based on whatever you have in the library, your past ratings, and titles with similar tags, the probability of showing different entries might change (it'll be considerably more expensive in terms of calculations, since it would no longer be a "static" (it currently gets updated on a weekly basis iirc) list for a given entry, and instead it'd be different based on the active user)

 

[Angry_Panda]

I don't have the specifics, but I'd assume it'll work in a similar way to how recommendations work in other platforms, where using the preferences from all the users, a map is generated where based on whatever you have in the library, your past ratings, and titles with similar tags, the probability of showing different entries might change (it'll be considerably more expensive in terms of calculations, since it would no longer be a "static" (it currently gets updated on a weekly basis iirc) list for a given entry, and instead it'd be different based on the active user)

EVERYONE ADD GAL☆CLEANING! TO YOUR LIBRARY RIGHT NOW.

 

[Bartolumiu]

EVERYONE ADD GAL☆CLEANING! TO YOUR LIBRARY RIGHT NOW.

Alright, time to blacklist that entry in the system then...

 

[BakedBanana]

EVERYONE ADD GAL☆CLEANING! TO YOUR LIBRARY RIGHT NOW.