resources-intelligibility

Some resources for intelligibility analysis of machine learning models, mostly in French.

Notes on setting up the project

• with a version of python3 installed (tested with python 3.6), make sure you have access to pip.

• with the below instructions, create a local virtual environnment and activate it

• install requirements.txt

  $ python3 -m venv .venv
  $ source .venv/bin/activate
  (.venv) $ pip install -r requirements.txt
  

• Go to the data/ folder and download (~0.5Mo) the required data with the link you can find in data/howtogetdata.txt. At the end of this step, you should have a carInsurance_train.csv file in the data/ folder.

• Start a jupyter server.

  (.venv) $ jupyter notebook
  

Features

In the notebooks/ folder, you will find some demos of several intelligibility techniques:

• Partie1_Construction_Modèle.ipynb
• Partie2_Analyse_sensibilité_des_prédictions.ipynb
• Partie3_Décomposition_en_contributions.ipynb
• Partie4_Décomposition_en_règles.ipynb

You should run Partie1 first because it will write a pickle with data and model, used by other notebooks. Afterwards, notebooks are independant.

Credits

This work has been done by Quantmetry R&D, 2018.