Some takeaways from CSCHW4 solution

readme.md

link to slides

Результаты:

Feature engineering	score (holdout)
no statistic	0.744318
basic statistic	0.764667
all statistic (basic + relative)	0.768469

Memory usage

• good intro
• smaller datatypes - some tips:
  • don't forget to bring nan back (or impute before)
  • change this: props.memory_usage().sum() / 1024**2 to this: props.memory_usage(deep=True).sum() / 1024**2
  • do something with object columns!

Feature importance

TL;DR use shap
why is it needed

• pretty good lightgbm kernel

Hyper parameter optimization

• docs
• general advice
• some code
• very comprehensive post

Rankinkg

• Catboost

  • tutorial
  • YetiRank paper
  • статья на хабре от победителей хакатона SNA 2019

• Lightgbm

  • example
  • docs
  • data format
  • usage from python

kaggle

• download a file without committing

lightgbm

• params, which I used
• pure python api kernel
• If you want to run lightgbm with GPU on Kaggle Kernel
  • Think once again :)
  • if still want - do like this