On machine learning model interpretability

Curated list of resources for data, science, and navigating life
exploratory-data-analysis
machine-learning
resources
chemical-science
Published

December 5, 2021

Process mindset vs outcome mindset argument – understanding right decision and right outcomes.

Helpful notebook on simple and useful tips on model interpretations in chemical science from ever-amazing Patrick Walters.

Few key work horses:

1. SHAP

SHAP (SHapley Additive exPlanations) is a game theoretic approach to explain the output of any machine learning model.Github

Considering cooperative prediction - the value added by the feature contribution to the final output and compare it to its individual output if that feature was active.

Given the current set of feature values, the contribution of a feature value to the difference between the actual prediction and the mean prediction is the estimated Shapley value.

2. Counterfactuals

If X had not occured, would have Y occured?

A counterfactual is a idea of relating an action to a consequence.

“Would I have got a cold, if I had not eaten the ice-cream?”

Usually a model agnostic approach is implemented, wherein the input(s) of the model is varied and its effect on the prediction is analyzed. Mind here that we dont really care how the model predicts tthe output but just if the output changes by changing the input.

The idea echoes with the concept of degree of rate control first proposed by Charles Campbell to propose kinetic pathways and intermediates which have most impact on the final chemical reaction rate.

Counterfactual have an important drawback - they suffer from the possibility of multiple truths. Explain on why that molecule: Andrew White Lab. Github

Pen’s blog on implementation of Exmol

ExplainerDashboard python package

Automate model design and NN architecture search?

3. Canonical Correlation Analysis

Video from Jay Alammar.