Sunday, May 28, 2006
Thursday, May 11, 2006
Every now and then the Economist publishes interesting graphs where they show how the count of the word 'Recession' in selected news publications correlates with the actual occurrence of an economic down turn. I always found this fascinating and used to think about all the cool stuff you could do if you had the actual counts of all the words in the all the news. And then I joined Google. Oh boy.
I remember talking to somebody from the news team and asking about this. Oh, well, we have that data around somewhere, the answer was. Using Googles Mapreduce it was actually quite simple to run over this data and extract the counts for words in the news over time. I threw together a quick http server that would actually plot the data and send it around the company. People thought it was pretty cool. Later I added the labeling of the peaks using the headlines of the news and some people started saying things like this should be launched on labs.
Then I met up with the guys doing something similar, but based on the number of times people searched for a certain keyword and we decided to combine the efforts. Oh and a designer had a look at the thing (my demo might have been cool but was also rather ugly). Yesterday the thing launched on Google Labs as Google Trends.
Thursday, May 4, 2006
The first time I heard the term Foosball it was during an episode of Friends. I thought it was some sort of joke, the Americans using the German term because it isn't really football and table-soccer sounds so odd. Maybe it is such a joke, but it seems it is the official term. Also, it seems to be a game popular among the Google going geeks. When I came to Google I fancied me to be somewhat of reasonable player. Man was I wrong. There's people doing trick shots I've only seen videos of. And then there's people with perfect control. One tricky thing is, how to create balanced teams (for playing two on two).
Actually, it started out with the desire to run some sort tournament. Problem is of course that if you pair anybody somewhat reasonable with Stefan, that team wins the tournament. What we need is a good ranking that spits out even pairs. Something like the ELO ranking.
So for a few weeks or so we tracked all games and wrote down the results. Last week I loaded up the results in a 50 line python program that does some machine learning. It actually predicts like 90% of all games correctly, which is not that far from optimal, given that the games are far from consistent with each other. And it probably is a simpler algorithm then the famed ELO.