After some preliminary research, I have noticed that people can successfully enter top 25% by forking some public kernel in the last week and enter top 10% by stacking several basic models. However, to achieve even better performance, people have to build a pipeline of a hierarchical model (usually 3-4 levels). I have collected some winners’ source code to learn from.

Competition Winner / Ranking Code Interview
Outbrain Team / 2nd Link Link
Santander Team / 2nd Link Link
Santander Ryuji Sakata / 3rd Link Link
CrowdFlower Chenglong Chen / 1st Link Link

Full list:

490 total views, 1 views today