On Counterfactuals and ML Explainability. Many businesses have some reluctance to adopting ML-based decisions. Usually, The Business Man believes that ML is a blackbox that is not explainable.
Counterfactuals (a.k.a. counterfactual explanations) can help in making blackbox models a bit more explainable by identifying what changes have to be done to modify the final ML model output/decision.
Counterfactual explanations describe how to modify a feature vector in order to flip the outcome of a trained classifier.
In Robust Counterfactual Explanations for Random Forests the team @PolytechniqueMontreal explain the relationship between the robustness of counterfactual explanations and the predictive importance of features.
Tree-Based Ensembles e.g. XGBoost are widely popular in financial ML. In Robust Counterfactual Explanations for Tree-Based Ensembles, the team @JPMorganAI Research introduce the concept of counterfactual stability; an attempt to quantify how robust a counterfactual is going to be upon model changes under retraining.
An issue with counterfactuals is dealing with real-world apps where categorical attributes have many different values or the ML model is non-differentiable. The team @Salesforce AI Research recently released a Model-Agnostic Counterfactual Explanation (MACE) Framework to address this issue.
There’s a lot of regulatory drive pushing on algorithmic fairness. The guys @TübingenUni, are focusing on counterfactuals and algorithmic recourse, as in providing recourse to individuals adversely impacted by model predictions. They’ve released CARLA - a Pythonic lib, notebooks, and datasets for counterfactual and recourse ML.
The “other” counterfactuals. A while ago, the great Ferenc [it’s been years since we organised a crazy ML hackathon together!] explains what are counterfactuals from a causal inference, probabilistic perspective, a.k.a. the Judea Pearl’s definition.
Have a nice Sunday.
10 Link-o-Troned
A Pythonista *Experience*
Scripting aRt
Love from Julia
data v-i-s-i-o-n-s
Distributed de-Entangler
Forschung!
Algorithmic Potpourri
Robots & Cyborgs like <you>
Deep & Other Learning Bits
startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.