notebooks

Hands-on Resources (ECML-PKDD 2020)

The Jupyter Notebooks included in this directory are hands-on materials used for the ECML-PKDD 2020 hands-on tutorial:

What and How of Machine Learning Transparency:
Building Bespoke Explainability Tools with Interoperable Algorithmic Components

These resources provide guidance on building bespoke surrogate explainers of individual black-box predictions for tabular data. See https://events.fat-forensics.org/2020_ecml-pkdd for more details.

Help with the Hands-on Exercises

During Part 3.2 of the tutorial you can reach out via the dedicated Slack workspace to ask questions and get help with any issues or problems. Registration for the workspace is limited to ECML-PKDD 2020 participants -- the sign-up URL can be found in the Whova app and on the ECML-PKDD 2020 website.

Contents

This directory holds five Jupyter Notebook, a Python library dedicated for this tutorial and a file with Python packages required to execute the notebooks.

• 0-environment-test.ipynb -- a test notebook used to validate environment setup.
• 1-data-sets.ipynb -- a notebook discussing two data sets that are used throughout the tutorial.
• 2-interpretable-representations.ipynb -- a notebook introducing interpretable representations.
• 3-data-sampling.ipynb -- a notebook covering various data sampling approaches.
• 4-explanation-generation.ipynb -- a notebook walking through explanation generation procedure.

---

• fatf_ecml.py -- a Python library with helper functions used throughout the notebooks.
• requirements.txt -- a list of (two) Python packages required to run the notebooks. (See below for more details.)

Executing the Notebooks

There are three distinct ways in which you can run the Jupyter Notebooks. First, you can install the required dependencies, download the notebooks and run them on your own machine. Alternatively, you can execute them online directly in your web browser (no installation required) using either of the two Jupyter Lab services: My Binder or Google Colab.

Local Installation

First you need to install the required Python packages. You can either do it manually with:

$ pip install fat-forensics[all]
$ pip install jupyterlab

or use the requirements.txt file:

$ pip install -r requirements.txt

Then you should run the Jupyter Lab server in the directory that holds the notebooks with:

$ jupyter lab

To download the notebooks (content of this directory) you need to get a copy of this GitHub repository. You can download a ZIP archive by licking this link. Alternatively, you can clone it with git using the following command:

$ git clone https://github.com/fat-forensics/Surrogates-Tutorial.git

Google Colab

To run the notebooks in your browser with Google Colab click the button below and select the desired notebook from the list (a Google account is required). Beware that any changes you make will not be saved unless you click the Copy to Drive button (visible at the top) and open your Google Drive copy of the notebook in a new tab when prompted.

My Binder

To run the notebooks in your browser with My Binder click the button below. Then, you need to navigate to the notebooks/ directory. Beware that any changes you make will not be saved unless you download the notebook that you are working on (use the File -> Download as -> Notebook menu option in Binder).