10th place solution to Sberbank Data Science Journey 2018: AutoML

SDSJ AutoML — AutoML(automatic machine learning) competition aimed at development of machine learning systems for processing banking datasets: transactions, time-series as well as classic table data from real banking operations. Processing is handled automatically by the system with models selection, architecture, hyper-parameters, etc.

Solution description

Based on @tyz910 's public kernel

Preprocessing:

All preprocessing procedures prallelized.

• If dataset's size bigger than 2Gb -> filtering columns with Boruta
• Drop constant columns
• Add is_na columns
• Filling na values and downcasting
• Extracting features from datetime columns
• Target encoding for string columns

Training:

While we have time:

1. Sample LightGBM hyperparameters and K-fold parameters(n_splits, shuffle)
2. Construct folds as subset from main train LightGBM dataset
3. Train all folds with LightGBM
4. Minimize oof-score with HyperOpt
5. Save all models

Local Validation

Public datasets for local validation: sdsj2018_automl_check_datasets.zip

• Official how to
• Baseline example
• @vlarine's public kernel.

Docker 🐳

docker pull ungvert/sdsj2018