leon-dae / explainable_ml_ice_study

Predicting compressive strength and behavior of ice and analyzing feature importance with explainable machine learning models

This repository is the official implementation of this paper. The codes were tested to work upon upload, but will not be further maintained.

Requirements

This research was done in an Anaconda environment, packages were installed with pip and conda. An environment.yml file can be found in the repository. The main required packages are

Name	Version	Build	Channel
xgboost	1.2.0-SNAPSHOT	pypi_0	pypi
tensorflow	2.2.0	pypi_0	pypi
pandas	1.0.3	py37h47e9c7a_0

Visualization packages are matplotlib and seaborn.

In order for the scripts to run, data_preprocessing.py and auxiliary_functions.py are required.

Corresponding data

The data required to train the models is in data_points.xlsx, version 1.12. This file is not available in the repository, but you can send an e-mail to [email protected] or [email protected] and we will share it with you.

Exploratory data analysis

The exploratory analysis is done with the following python codes: data_preprocessing.py, exploratory_all_data.py and exploratory_strength_values.py.

Training

To train the models in the paper, run the following notebooks: behavior_XGBoost.ipynb, strength_XGBoost.ipynb, behavior_ANN.ipnyb, strength_ANN.ipynb.

Comparison models

Different approaches were compared to the machine-learning models. Regarding the strength regression ML models, the comparison models were empirical formulas. Their computation can be found in strength_empirical.py. Regarding the behavior classification ML models, the analytical comparison model is in behavior_analytical.py.

Evaluation

The results were evaluated within the model training notebooks and in model_performance.py regarding taking the log of the target strength values or not.

Pre-trained Models

Trained models are available in pickle format from the repository. They can be run with

model_path_name_template = 'models/rgsr_xgb_sw_pickle.dat'
model = pickle.load(open(model_path_name_template, "rb"))  
predictions = model.predict(xgboost.DMatrix(X))

Contributing

If you found any mistakes, or would like to contribute to the models, please drop me a message: [email protected].