LightGBM and basic features

chestnut 🌰

🌰 code

Run experiment!

Random search in the hyper-parameters space

• Make sure that you put correct paths to data: neptune_random_search.yaml:L21. Consider your own ranges of hyper-parameters and number of runs.
• neptune run --config neptune_random_search.yaml main.py train_evaluate_predict --pipeline_name lightGBM

Single run

• Make sure that you put correct paths to data: neptune.yaml:L21
• neptune run --config neptune.yaml main.py train_evaluate_predict --pipeline_name lightGBM

In both cases pipeline is called lightGBM

Solution Overview

Features

First solution uses some of the features from application_{train, test}.csv provided explicitly in the pipeline_config.py, that is CATEGORICAL_COLUMNS and NUMERICAL_COLUMNS.

Model

The model is LightGBM and is defined here: models.py:L7. Model accepts parameters from the neptune.yaml:L39 in case of the single run or from the neptune_random_search.yaml:L39 in case of the random search

Postprocessing

We do simple predictions clipping, defined in the post postprocessing.py:L6