Skip to content

Automated approach from feature engineering to modeling on the Kaggle Home Credit Default Risk competition dataset

Notifications You must be signed in to change notification settings

aldente0630/loan-repayment-prediction

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Automated Feature Engineering with FeatureTools

Relationship Diagram

  • A dataset with 2,221 features for 356,255 customers was finally created. Saved as a CSV file, it is about 4GB.
  • The whole process took 3 hours and a half on my iMac with 6 cores and 16GB of memory.

Automated Modeling with AutoGluon

  • The goal is to do automated modeling with AutoGluon.
  • AutoGluon makes it easy to automatically experiment with a variety of algorithms, from tree ensembles to deep learning and even model stacking.
model score_val pred_time_val fit_time
weighted_ensemble_k0_l2 0.787430 3098.737486 95757.520068
weighted_ensemble_k0_l1 0.786499 601.329862 46704.633752
CatboostClassifier_STACKER_l1 0.786261 2511.553999 53790.340020
LightGBMClassifierXT_STACKER_l1 0.785994 2511.152501 53834.121477
LightGBMClassifier_STACKER_l1 0.785990 2511.691034 53782.292310
LightGBMClassifierCustom_STACKER_l1 0.785596 2510.629085 54090.092252
LightGBMClassifierCustom_STACKER_l0 0.782958 10.546562 1941.627757
CatboostClassifier_STACKER_l0 0.782336 7.888541 1890.238214
LightGBMClassifierXT_STACKER_l0 0.780601 11.507542 860.345474
LightGBMClassifier_STACKER_l0 0.780356 10.297791 824.519218
...
  • The model stacking technique achieved the highest predictive performance. This was 0.78149 for the Kaggle public board and 0.78391 for the private board as measured by AUROC.
  • This process took about 1 day and 6 hours to train on an AWS m4.16xlarge EC2 instance with 64 cores and 256GB of memory, and about an hour and a half to infer.

About

Automated approach from feature engineering to modeling on the Kaggle Home Credit Default Risk competition dataset

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published