hyperpape, all I am arguing is that we need more data to make more confident assumptions. Lee Sedol and Iyama Yuna haven't played one another much, so we don't know much about what their relative placement should be.

In the same way, Iyama Yuta hasn't played a lot of people on the list, so it's hard to know what his actual placement should be.

The set of people that have played each other more often - those guys - we know more about their relative placement.

End of story.

_________________
be immersed

Also, I should clarify something.

I am not saying that it is impossible to project an assumption about Iyama's position internationally. Goratings.org has done that, and in my opinion, the placement is probably pretty good. But I do believe that more data among players in the population makes the placement assumption more accurate. If Iyama Yuta keeps playing those nearby him in the list, we can have more confidence in where he should be. So naturally, players that have played one another more on this list - we can be more confident of our assumptions about those guys, because we have more data.

It would be nice to quantify our certainty in some way of the ratings. So for players that have played most of the people on the list many times, we'd be able to see more confidence in our assumption about their placement. Whereas if you have a player that has only played 2 games, we could express our confidence in that guy's placement as being weak (e.g. maybe he got lucky).

_________________
be immersed

I agree with this post, though I would reframe it a bit:

There's already a concept of confidence in the model and I believe Remi tweaked the display to show the lower bound of the confidence interval: viewtopic.php?p=190653#p190653.

With Iyama, I suspect that the model has normal-high confidence in his rating, since he's played many games. We're all standing around speculating as to whether the model reliably/accurately reflects reality.

_________________
Occupy Babel!

I guess the only part where my opinion differs is that I think that high confidence should come from the rating if he has played many games against many different people from the population that he is being ranked in.

To give an extreme example, if he only played one guy on the list a million times, I'd personally have little confidence in whatever ranking we give him, even if he won (or lost) all of the time against that guy.

_________________
be immersed

Last month I count the nation of pro in goratings
Japan 333
China 251
South Korea 193
Taiwan 28

So maybe this is the problem, since Iyama Yuta play with the biggest group of the population (Japan), it'll continue to give him a good record
But the model will still give a good prediction, since it's more likely that he'll continue to play a lot of domestic games. If you lower the Elo of Iyama Yuta, it might reflect a better prediction for game against Korean/Chinese pro (which he rarely play) and worse prediction for domestic game (which he'll play a lot)

I agree. It would almost make sense to calculate rating in a different way, depending on the page you are viewing. The ranking/results you are aggregating should reflect the dataset it was calculated from.

I also agree that the model can have good prediction, given that the dataset Remi uses is representative (and continues to be representative) of all of the games that pros play. If the dataset doesn't include all games we know about, it could be skewed.

_________________
be immersed