There is an increasing need for transparency and fairness in Machine Learning (ML) models predictions. Consider for example a banker who has to explain to a client why his/her loan application is rejected, or a health professional who must explain what constitutes his/her diagnosis. Some ML models are indeed very accurate, but are considered hard to explain, relatively to popular linear models.
Source of figure: James, Gareth, et al. An introduction to statistical learning. Vol. 112. New York: springer, 2013.
We do not want to sacrifice this high accuracy to explainability. Hence: ML explainability. There are a lot of ML explainability tools out there, in the wild for that purpose (don’t take my word for it).
The teller is a model-agnostic tool for ML explainability - agnostic, as long as this model possesses methods fit and predict. The teller’s philosophy is to rely on Taylor series to explain ML models predictions: a little increase in model’s explanatory variables + a little decrease, and we can obtain approximate sensitivities of its predictions to changes in these explanatory variables.
Installation
Currently from Github, for the development version:
pip install git+https://github.com/Techtonique/teller.git
Package description
This notebook will give you a good introduction:
thierrymoudiki_011119_boston_housing.ipynb
Two models are used in the notebook: a linear model and a Random Forest (here, the black-box model). The most straightforward way to illustrate the teller is to use a linear model. In this case, the effects of model covariates on the response can be directly related to the linear model’s coefficients. Also, note that if there a lot of variables in your model, the teller’s explainer can be created with option n_jobs=-1 (for parallel execution).
Contributions/remarks are welcome as usual, you can submit a pull request on Github.
Note: I am currently looking for a gig. You can hire me on Malt or send me an email: thierry dot moudiki at pm dot me. I can do descriptive statistics, data preparation, feature engineering, model calibration, training and validation, and model outputs’ interpretation. I am fluent in Python, R, SQL, Microsoft Excel, Visual Basic (among others) and French. My résumé? Here!