| Yeah, zarin said it a bit better than I did. Well put. |
| #dontcare |
aisakku said: Kei: what did you think of further normalization? hmm? |
| heh. thanks for the clarifications. I guess for the majority of users the total mean might be better. Even for me, it'll still be a good measure of compatibility. Over time, it'll only get even more accurate. |
kei-clone said: aisakku said: Kei: what did you think of further normalization? hmm? aisakku said: Edit 2: for anyone curious about a further normalized system: |
| #dontcare |
| I like that normalization in that the score spread my vary more, but I don't believe formula is correct for usera < meana For "D(a) = usera(usera-meana)/(meana-1)," let's say mean = 8, that would mean that both a score of 1 and a score of 7 gives the same answer -1 as the value relating to the score below their mean. |
aisakku said: kei-clone said: aisakku said: Kei: what did you think of further normalization? hmm? aisakku said: Edit 2: for anyone curious about a further normalized system: hmm...I see the intention here but I'm not sure how much benefit this will bring over just straight up taking the difference between the mean and the score. I can see an immediate result though that all the numbers will result in a lot closer to a zero value though, and not sure that's a good thing. |
kei-clone said: aisakku said: kei-clone said: aisakku said: Kei: what did you think of further normalization? hmm? aisakku said: Edit 2: for anyone curious about a further normalized system: hmm...I see the intention here but I'm not sure how much benefit this will bring over just straight up taking the difference between the mean and the score. I can see an immediate result though that all the numbers will result in a lot closer to a zero value though, and not sure that's a good thing. Yeah that is true, plus I'm pretty sure xinil would not want to code all of that. scorpedo said: I like that normalization in that the score spread my vary more, but I don't believe formula is correct for usera < meana For "D(a) = usera(usera-meana)/(meana-1)," let's say mean = 8, that would mean that both a score of 1 and a score of 7 gives the same answer -1 as the value relating to the score below their mean. Ahh yeah i see, though that'd be fixed if you change it and use the lowest output score instead of 1 for the denominator. Edit: Oh I see the flaw with it.... would have to change the formula to be something like: (meana-usera)(usera-meana)/(meana-alow) for meana > usera (making these formulae up from scratch so thanks for pointing that out.) That would also turn usera > meana into: (usera-meana)(usera - meana)/(10-usera) |
aisakkuJan 26, 2008 11:32 AM
| #dontcare |
| I just realized that at the 'grand totals' on each list, Xinil already displays each users deviation based on the average of all members. Can that somehow be used instead of the user's mean? Just throwing that out there in case it's easier/more accurate that way; haven't put too much thought into it. |
scorpedo said: I just realized that at the 'grand totals' on each list, Xinil already displays each users deviation based on the average of all members. Can that somehow be used instead of the user's mean? Just throwing that out there in case it's easier/more accurate that way; haven't put too much thought into it. I thought about that before too. But never was able to come up with a formula. |
Zarin said: Here's some random comparisons between me and other MAL users: Looking through it there really isn't a huge variety in similarities, which is a bad thing as it's kinda hard to look at the number and now exactly how similar you are. Adding text may help but with that small of difference it will be hard to properly correlate correct words to the scores. Again, the numbers for mine are currently off and should not be trusted. Both systems are going to differentiate from the original, as well as from each other, The output from my system would be read the same as the previous, with 0 being the best possible score and 9 being the worst. It's possible to convert mine into percentages as well, but for now I believe we're mostly looking for input |
| #dontcare |
| I think, unless there's a large amount of disagreement, I'm just going to go with abh's percentage correlation. Sure, it's not as great as aisakku's, but it's easier to implement, and offers a better 'visual' and 'immediate' representation. If you guys disagree and want to get a poll going, go for it. I just want some movement on this. |
| sounds good to me |
| My only real disagreements have been stated. His system is better than the current, so go for it. |
| #dontcare |
| aye, abh's system is fine with me. |
| No objections. |
 staff.applications ▼ guidelines.faq ▼ report.abuse ▼ thx.skittles ▼ thx.kina ▼ [H+] ³ ▼ |
| As noted in the other thread, Xinil, I recommend putting a higher limit on the number of entires needed to have a compatibility rating done (or one in general if there isn't any right now). Especially with a system like this, it's going to be very inaccurate with only a couple of entires. |
| #dontcare |
| First, thanks for implementing my system. As aisakku said there should be at least a certain fixed number of entries for good accuracy. It is also so in my userscript. And thanks to all the people who contributed to this discussion over past few months. Now I am going to work upon a new proposal, an API for MAL, which will enable developing applications for MAL, such as a tracker. |
  |
| Does the greasemonkey script currently use aisakku's method? I was reading in detail until somewhere in page two, at which point I started skimming. Lots of math and I got a bit lazy. Anyway, the gist I got from it is that aisakku's method detects the similarity of scoring patterns between users, which would account for differences in personal rating scale (for example, 6 is "true neutral" for me, despite 5 being labeled "average", which I consider to be mildly negative). If the script uses said method, I would prefer that to the current system. I'm a bit saddened a poll was never started here (although I wouldn't have been able to participate in it at the time). I feel rating patterns would make for a much better comparison than simple vector comparison, since vector comparison is still tied to the "default" rating system, which many users (including myself) may not abide by. |
TripleSRank said: Does the greasemonkey script currently use aisakku's method? Yes. And the current system adopts that method (i.e. Pearson product-moment correlation coefficient), which is resistant to simple linear transformation (e.g. ax+b). |
 |
RitaDnz said: abhin4v said we could download the scrip from http://userscripts.org/scripts/show/12746 but that url seems to be broken. Where can I get it now? http://userscripts-mirror.org/scripts/show/12746 However, I think Xinil already implemented it natively into the site, so I am not sure you need it |
   |
More topics from this board
» To Make Manga List Updating More ConvenientSaintseiya100 - 2 hours ago |
0 |
by Saintseiya100
»»
2 hours ago |
|
» An option to see which users have favorited a series or charactersJKKH - Oct 7 |
2 |
by Shishio-kun
»»
Yesterday, 1:07 PM |
|
» new list option - "Maybe"dailydi - Oct 12 |
8 |
by Shishio-kun
»»
Yesterday, 7:25 AM |
|
» Yearly Wrapped like MyDramaList ?ame - Sep 6 |
13 |
by ScaryOwl
»»
Oct 11, 10:47 AM |
|
» Abusing the Number of Shared Anime and Affinity to Uncover Private Listsvazae - Mar 7, 2021 |
6 |
by Serhiyko
»»
Oct 10, 7:01 AM |