This is more curiosity than anything else, but what’s going on with 本好き’s difficulty rating? It was at 31 basically forever, then dropped to 30 for a few weeks, then went back to 31, and now is at 32. I thought the ratings weren’t supposed to be this volatile, particularly for books with many ratings?
4 Likes
No, remember that ELO doesn’t make the ratings stickier as there becomes more gradings… as it assumes the entire notion of a lvl 31 or lvl 30 may be changing.
As we’ve discussed before, we could try other rating systems, but really focusing on movies right now… which has been more difficult than I imagined, as I’m trying to make the entire system more flexible for additional future types 
5 Likes
I do think the ELO system is a bit strange for book rating, since (as I understand) it’s a floating value, which gives more immediate weight to new gradings. This makes sense in chess or another competitive game, where players can get better or worse over time, but for books it would make more sense for old ratings to have equal weight with new ones.
Ideally, I think it would also be good to have an algorithm that checks the ‘transitivity’ of ratings too, so to speak. By that I mean if Book A has consistently been rated harder than Book B, then when Book B’s difficulty level increases, Book A should automatically increase accordingly. I’m not sure an efficient algorithm like this even exists, though… Sounds like complicated graph theory lol.
8 Likes
Aren’t the ELO Ratings more designed towards skill of the user, instead of skill required for a medium?
Also how does it fare being media instead game skills such as chess?
In a game, you start with a white canvas and then go from there.
Unless there is stuff like puzzles that are based on ELO scores?
2 Likes
Sounds like some kind of weighted, directed graph. Except with the added complexity of resolving conflicts (i.e. when transitivity is contradicted by the actual gradings). I haven’t done much with graph theory in nearly a decade, so I don’t have anything useful to add beyond that. 
It would be kind of fun if @brandon publicly shared the gradings dataset for people to experiment with.
2 Likes
You are correct that there are ‘static’ Elo rating system (see Bradley-Terry model) which have to be recalculated for every book every night, which i’ve played around with briefly. However, it has issues as well - sometimes you get books in rating islands not connected to the larger rating system and those books’ ratings don’t resolve to a value.
Yes that’s true. However, there’s no rating system tailor-made for the present application, so we do have look at what’s out there and use whats best 
.
I would be happy to share a partial dataset if people were actually interested in playing around with it. Just ping me
I will say though that while the accuracy of the Elo system may be improved, it has the end functionality I want a rating system to have. Namely:
- it generates a rating value
- it accepts pairwise comparisons
- it can handle rating islands & temporary ratings relatively well
- easy to calculate, easy to understand, easy to diagnose issues
Potential improvements:
- static ranking, rather than dynamic ranking (see Bradley Terry)
- ability to generate personal ranked lists
- ability to allow users to grade all their comparisons, rather than limited to 6 per book (elo would give them too much power)
- stickiness of heavily graded books (would be covered by static ranking)
As I noted, Elo is very simple to use and has the end functionality that I want, which is why its stuck around as long as it has. Once I finish the major additions I’m planning (movies / tv shows, a second language, high priority platform features - tagging, descriptions… a few others), I thought it would make some sense to dedicate time to the grading system.
Elo works surprisingly well though - it just has a tendency to oscillate up and down for highly rated books such as @seanblue mentioned. There are just lots of things to do
8 Likes
The exact opposite of the expected behavior of a rating system I think. 
Well hopefully someone responds here about how they have a PhD in graph theory and want to help. 
2 Likes
You can study some data analyst / data scientist PhD books in Japanese, and that’s two birds with one stone! 
Thanks for the detailed explanation.
Maybe there’s a way to hold highly oscillating books without changing the algorithm if it’s already producing satisfying results?
I did spend several days last year reading Japanese Wikipedia pages on math. This seems like the natural next step!
4 Likes
Here’s one that is not copyrighted (Creative Commons license, see page 3)
2 Likes
Elo has some limitations due to being designed to be able to be calculated by hand in a time when computers were less common place.
Most chess sites etc, use improvements like Glicko-2 instead. Which to my knowledge adds:
- ratings becoming more stable the more comparisons are added.
- ratings slowly becoming more volatile again over time if no new comparisons are added.
Although I don’t know whether the latter bit makes sense for Natively. (I don’t think large amounts of Elo drift over time are particularly common.)
Might be worth looking into.
5 Likes
Hey @brandon, per my comment here, I no longer think 僕が愛したすべての君へ | L30 and 君を愛したひとりの僕へ | L30 should be listed as a series. I think they should be split into separate entries, maybe with a note saying they are related to each other. This would match Amazon and AniList (for the upcoming anime adaptions), which also list them individually. Now that I’ve finished both books, I can say for sure that the order I originally suggested is definitely not the “right” order to read them in. If either is “right”, it’s the opposite, simply based on the chapter structure. But in the end I think they are more like parallel stories that can be enjoyed in either order, and which order is better will probably depend on the person.
4 Likes
Noted and thanks!
I’ve now split them (1, 2) and erased the first book gradings from affecting the second book. I’ve set the second book at a temporary lvl 30 
I’m not too worried about showing the relations if it’s simply the same universe - they are easily viewable if you search by author.
3 Likes
Well, at the very least they are from the same multiverse. 
I’ll try to add reviews to both of them, so I’ll mention that they are related there.
2 Likes
They still appear as a series in my library. Will that fix itself at some point, or do you have to fix that manually?
EDIT: It’s fixed now
EDIT 2: But I can’t grade the second book for some reason.
1 Like
Specific to this forum (Discourse) can I ask that allow user locale and set locale from accept language header be enabled? It does not currently seem to be. Basically I’d like the languages Discourse is already available in to be auto-translated into several languages and it looks like there is functionality to automatically pick up user language (here - scroll to bottom it’s a long thread).
It looks like this when enabled, fwiw:
6 Likes
Tagging you @brandon, in can you didn’t notice this.
1 Like
Sorry about that - should appear now! When I broke them apart I forgot to do something for your personal grading data (you needed the your grading status of that book updated), thanks for catching.
Ok updated!
Apologies I didn’t get back to these messages until today, busy weekend
7 Likes
Funny side effect
5 Likes
Interesting. I see レギュラー for you and Regular for me. 
I’m surprised your (?) language settings affect my display.
4 Likes
