README.md

Kaggle-Home-Credit-Default-Risk

This repo aims to kaggling Home Credit Default Risk in a pipeline fashion. This pipeline is not necessary a performer (PB 0.792/ LB 0.796, top 25~30%) but, aiming to bring a bit automation and using config files to control its process. It does, 1) data set construction ann caching, 2) feature transform, 3) hyperpparameters search and 4) stacking models to generate submission file.

The necessary data set and descriptions can be fould here. https://www.kaggle.com/c/home-credit-default-risk

Modules

ModelPipeline.py is exposed to execute all the tasks and the other main modules are located in lib. In lib, DataProvider.py constructs data set and use FeatureTransformer.py to do feature engineering. ScikitOptimize.py use Bayesian Optimization in Scikit-Optimize to perform hyperparameter search. AutoStacker.py uses mlxtend to build stacking classifier. The shared variables and templates are kept in LibConfigs.py. Some small tools are implemented in Utility.py. DataFileIO.py is responsible for saving and loading HDF5 and CSV files.

Folders

The cached data are stored in data, found hyperparameters are stored in params, result for submission in output.

Configs

In configs, there are three configs to control the whole modeling process:

configs for feature generations -- SampleDataConfigs.py

configs for hyper parameter search -- SampleModelConfigs.py

configs for feature stacking model -- SampleStackerConfigs.py

Usage

Load data set and create caches for train and test dataframes

python3 ModelPipeline.py -a cache_prefix

Re-construct train and test data set, use after modified *DataConfigs

python3 ModelPipeline.py --refresh-cache

Do hyperparamters tuning on base model

python3 ModelPipeline.py --compute-hpo -t LGBM,LossGuideXGB

Before doing stacking, adding some external results from forked kernels (optional)

python3 LGBMSelectedFeatures.py

Prepare meta features for stacking

python3 ModelPipeline.py --compute-stack --refresh-meta

Introduce OOF results from external models (optional)

adding filenames such as probs_*.hdf5 into variable ExternalMetaConfigs in the using *StackerConfigs.py

Last step, do stacking

python3 ModelPipeline.py --compute-stack

Other flags

--debug --enable-gpu