Idea
===
Blending.
Linear blending is not enough to get a high score.
I split the test data into several data set and blend each 
with different predictors and different weights.

Language , children and grandparent predictors are not so usefull
in my solution. 

I don't figure out  why filters didn't work neither, oooops.

The idea blending method should hits 2800 in my test data, while 
only nearly 2400 were hit actually. 

Apcluster produce lots of cluster and doesn't work in this solution, if I choose middle value as preference.



Todo
====
Learn the weights in KNN
More good clusters
More smart blending method
Read the codes and idea of the leading guy carefully.


Credits
=======
Thanks to:
Daniel Haran && xlvector && jeremybarnes who contribute the ideas and basic codes.