|
On March 23 2014 06:43 ZERG_RUSSIAN wrote:
Is there any way to control for Elo in these? Like if you only take diamond 5 or above in solo queue?
That would be the shit and probably far closer to theoretically sound. And Ahri should get dunked by a good Lux imo, even if the data doesn't show it.
Unfortunately, no. Eiii did not capture leagues, or whether or not it's ranked (yes, it could even be normals).
Right now it's more on developing the model so I can analyze this kind of data effortlessly. As far as actually mining the data, it's time consuming - but otherwise straight forward... just need to use the Riot API.
|
United States23745 Posts
I'm pretty sure he said it would be pretty easy to mine for that data, he just hand't been doing it.
|
I read mime the data
that would be impressive
|
On March 23 2014 04:31 Sufficiency wrote:Show nested quote +On March 23 2014 03:57 Ketara wrote:
Are the reds the ones she does well against or the ones she does poorly against? I'm pretty sure blue is good and red is bad yes?
And these are adjusted for said champions overall winrate?
I really don't understand what the P value means. Blue -> Lux is strong against Red -> Lux is weak against. Yes, adjustments were made already. EDIT: read the table as follows. p-value < 0.02: strong evidence of good/match match up p-value > 0.02 but < 0.05: weak evidence p-value > 0.05: no evidence
Its important to note two things
1) There is almost guaranteed to be omitted variable bias in this because interaction terms make the calculation significantly difficult enough to not do it and because we expect that strong interactions are played more often than weak interactions. I wouldn't worry too much about it(mainly due to calculation difficulties) but know its there.
2) That probably isn't the p-value means in this sense. You should be able to notice that there is strong correlation between the p-value and the absolute effect. Its probable that the expected standard deviation for each variable is roughly the same size. If that is the case, our evidence for a specific value isn't any better or worse its just that for the variables for which are close to zero our estimates are close to zero.
The p-value issue is exacerbated because we have so many variables. Supposing the data was actually random we would expect about 5 values in such a test to have lower than .05 p-value.
As an aside, what precisely is the model you're running?
|
Graves needs more damage in general though. It doesn't matter that his burst in the midgame is slightly better if Caitlyn, Ezreal, etc. deal more damage than you (even assuming Ezreal doesn't use his passive) because they have more range and get to attack much more. Either you make him resilient enough that he can stick that close longer and loses less damage that way, or you make his dps scale better in the late game so his Q doesn't just tickle bruisers and then he runs.
(That or they could make his midgame stronger, but that's not Riot's thing.)
|
On March 23 2014 07:08 onlywonderboy wrote:
I'm pretty sure he said it would be pretty easy to mine for that data, he just hand't been doing it.
correct. There's the minor problem where you can't tell what 'skill level' individual games are played at, just the skill levels of all of the players in the game when you get the data-- which might not be the same as the skill level they were at when they played the game.
Shouldn't really matter though.
EDIT: btw if anyone can get me the riot hookup for an unlocked API key this could go way faster, but I don't know how much they like to give those out.
|
Can someone tell me what's going on with EG in this LCS game? A whole new line-up ~_~
Edit: nevermind, found the answer in the tourney thread. Visas too strong.
|
On March 23 2014 07:41 Cedstick wrote:
Can someone tell me what's going on with EG in this LCS game? A whole new line-up ~_~
Snoopeh/Krepo/Yolopete need to get their visas redone much like Bjerg did in previous weeks so the 2 remaining EG players got some subs together.
|
Don't even notice the vi nerfs at all
|
people build damage on her so she actually got buffed
|
On March 23 2014 07:07 Sufficiency wrote:Show nested quote +On March 23 2014 06:43 ZERG_RUSSIAN wrote:
Is there any way to control for Elo in these? Like if you only take diamond 5 or above in solo queue?
That would be the shit and probably far closer to theoretically sound. And Ahri should get dunked by a good Lux imo, even if the data doesn't show it. Unfortunately, no. Eiii did not capture leagues, or whether or not it's ranked (yes, it could even be normals). Right now it's more on developing the model so I can analyze this kind of data effortlessly. As far as actually mining the data, it's time consuming - but otherwise straight forward... just need to use the Riot API.
Do you have data for allied champions?
Like, Lux wins games X more often when X champion is also on her own team?
|
On March 23 2014 07:08 onlywonderboy wrote:
I'm pretty sure he said it would be pretty easy to mine for that data, he just hand't been doing it.
Yes, exactly. It's on the output Riot gives you - just need to record it.
On March 23 2014 08:15 Ketara wrote:Show nested quote +On March 23 2014 07:07 Sufficiency wrote:On March 23 2014 06:43 ZERG_RUSSIAN wrote:
Is there any way to control for Elo in these? Like if you only take diamond 5 or above in solo queue?
That would be the shit and probably far closer to theoretically sound. And Ahri should get dunked by a good Lux imo, even if the data doesn't show it. Unfortunately, no. Eiii did not capture leagues, or whether or not it's ranked (yes, it could even be normals). Right now it's more on developing the model so I can analyze this kind of data effortlessly. As far as actually mining the data, it's time consuming - but otherwise straight forward... just need to use the Riot API. Do you have data for allied champions? Like, Lux wins games X more often when X champion is also on her own team?
Definitely possible.
|
On March 23 2014 07:57 kongoline wrote:
people build damage on her so she actually got buffed
Yeah people routinely build more than 200 bonus AD on her.
|
Gah I think I made a mistake with my previous analysis. But now I am getting these CRAZY results and I am not sure of anything anymore, lol.
Epic fail from my part.
EDIT: OK found some more mistakes. This is going to need some major overhaul.
|
On March 23 2014 08:53 Sufficiency wrote:
Gah I think I made a mistake with my previous analysis. But now I am getting these CRAZY results and I am not sure of anything anymore, lol.
Epic fail from my part.
What is your model?
|
On March 23 2014 08:54 Goumindong wrote:Show nested quote +On March 23 2014 08:53 Sufficiency wrote:
Gah I think I made a mistake with my previous analysis. But now I am getting these CRAZY results and I am not sure of anything anymore, lol.
Epic fail from my part. What is your model?
It's just logistic regression, but I need to transform the data correctly, which I phailed at.
The bigger problem is what if the two champions in question are on the same team. The results I got so far I just shoved that problem under the rug, but I am considering a more reasonable solution.
At this point it seems I should just analyze synergy first (like what Ketera suggested) instead of analyze counters.
|
On March 23 2014 09:04 Sufficiency wrote:Show nested quote +On March 23 2014 08:54 Goumindong wrote:On March 23 2014 08:53 Sufficiency wrote:
Gah I think I made a mistake with my previous analysis. But now I am getting these CRAZY results and I am not sure of anything anymore, lol.
Epic fail from my part. What is your model? It's just logistic regression, but I need to transform the data correctly, which I phailed at. The bigger problem is what if the two champions in question are on the same team. The results I got so far I just shoved that problem under the rug, but I am considering a more reasonable solution. At this point it seems I should just analyze synergy first (like what Ketera suggested) instead of analyze counters.
Your data is binary is it not?
edit: But really that doesn't tell me much about your model. How are you controlling for the effect of a champions win rate (that you aren't looking at?)
|
On March 23 2014 09:21 Goumindong wrote:Show nested quote +On March 23 2014 09:04 Sufficiency wrote:On March 23 2014 08:54 Goumindong wrote:On March 23 2014 08:53 Sufficiency wrote:
Gah I think I made a mistake with my previous analysis. But now I am getting these CRAZY results and I am not sure of anything anymore, lol.
Epic fail from my part. What is your model? It's just logistic regression, but I need to transform the data correctly, which I phailed at. The bigger problem is what if the two champions in question are on the same team. The results I got so far I just shoved that problem under the rug, but I am considering a more reasonable solution. At this point it seems I should just analyze synergy first (like what Ketera suggested) instead of analyze counters. Your data is binary is it not? edit: But really that doesn't tell me much about your model. How are you controlling for the effect of a champions win rate (that you aren't looking at?)
By not looking at them, of course! 
|
does anyone think a double lifesteal with last whisper build path for an adc would work in this meta? Basically just stay alive and do more damage than the other adc overall in fights, and live longer through sheer lifesteal? Thinking about trying out a BT -> LW -> Bork build, and then adding in an IE and defensive item after that. Thoughts?
|
On March 23 2014 07:15 Goumindong wrote:Show nested quote +On March 23 2014 04:31 Sufficiency wrote:On March 23 2014 03:57 Ketara wrote:
Are the reds the ones she does well against or the ones she does poorly against? I'm pretty sure blue is good and red is bad yes?
And these are adjusted for said champions overall winrate?
I really don't understand what the P value means. Blue -> Lux is strong against Red -> Lux is weak against. Yes, adjustments were made already. EDIT: read the table as follows. p-value < 0.02: strong evidence of good/match match up p-value > 0.02 but < 0.05: weak evidence p-value > 0.05: no evidence Its important to note two things 1) There is almost guaranteed to be omitted variable bias in this because interaction terms make the calculation significantly difficult enough to not do it and because we expect that strong interactions are played more often than weak interactions. I wouldn't worry too much about it(mainly due to calculation difficulties) but know its there. 2) That probably isn't the p-value means in this sense. You should be able to notice that there is strong correlation between the p-value and the absolute effect. Its probable that the expected standard deviation for each variable is roughly the same size. If that is the case, our evidence for a specific value isn't any better or worse its just that for the variables for which are close to zero our estimates are close to zero. The p-value issue is exacerbated because we have so many variables. Supposing the data was actually random we would expect about 5 values in such a test to have lower than .05 p-value. As an aside, what precisely is the model you're running?
You are making this way too complicated than it is. I lowered the cut-off to reduce FDR, that's it. For each champion match up the model is plotted separately.
|
|
|
|
|
|
EPT
