Robert_Charlton - 6:35 pm on Sep 25, 2012 (gmt 0)

Here's a thought. What if G was "training" their filters instead of assigning them a value. Take Geo-location. Instead of using Server IP, ccTLD, potentially whois and WMT settings- what if they just let "the world" tell them which locations responded well to a site, and which ones don't

Shaddows - This is the concept I've been circling around in a bunch of earlier posts, but I've never articulated it nearly so well. I've been calling it "calibration", which is very fuzzy thinking. "Training" crystallizes the idea for me.

Regarding tedster's comment...

...the machine-learning does seem to be going pretty slow in the mid to long tail queries.

It would necessarily be going slower, because it would have less data on less frequently searched queries.