far0n / kaggle-homesite Goto Github PK

Hi @Far0n

First i would like to thank you for sharing this script and congratulation for your Kaggle competition !

I try to reproduce the process of feature selection with your xgbfi.

As I understand the process :

• 1. Fit a xgboost with original train file. I dump the model and I can analyse the feature importance with xgbfi.
• 2. Keep only the best feature (your top111) and find interaction between features (interactions2way, interactions3way...)
• 3. Reproduce process 1) and use CV to tune model, maybe add or delete some others features.

Is it the good way to go ?

I spent many hours and night to try it, but I don't find the same result at all. For example during the 1) process, I find PropertyField37 very effective for 2 Interactions features. But it seems you didn't choose it. Can't understand why.

In the same case I have only 100 features used by xgboost. Some are present in your top111 and some are missing. So much frustration !

If you are allowed to help me, you will save me some nights.

Got two questions regarding the code in top15_homesite_xgb.py:

1. Why separate x_train into dtrain and dtrain_tc and trained with two stages?
2. Why 300 rounds and eta = 0.01 for dtrain_tc training? Were those from CV results?