Source: Matt Dancho (Business Science)
@mdancho84
Explaining black box machine learning models is critical to gaining leadership’s buy-in and trust. Here’s 6 months of research on Explainable ML in 6 minutes (Business Case included). Let’s go!
1. Explainable Machine Learning (ML): Refers to techniques that make the outputs and operations of machine learning models understandable to humans. Traditional machine learning models, especially complex ones like deep neural networks, are often seen as “black boxes” because their internal workings are not easily interpretable.
2. Black-Box Problem: People don’t trust what they don’t understand. It’s that simple. With Explainable ML, you gain: Transparency, Interpretability, Accountability, Fairness and Bias Detection, and Trust. This builds confidence among stakeholders, which is especially important in domains like marketing, finance, and healthcare.
3. The 2 Types of Explainability Approaches: Model-Specific and Model Agnostic. Let’s break them down.
4. Model-Specific Explainability: Some models are explainable without any added processing. These tend to be simpler models. Linear Regression Coefficients: In linear models, the coefficients indicate the importance and direction of the influence of each feature. Decision Tree Rules: Decision trees provide a clear set of rules and thresholds for decision-making, making them inherently interpretable.
5. Model-Agnostic Explainability: These are methods that can be applied to ANY model. Examples include Feature Importance Scores, SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Partial Dependence Plots (PDP).
6. Why I use Explainable ML? In my $15,000,000 lead scoring model, it initially started out as a Linear/Logistic Regression. This was a simple model. But eventually I upgraded. I went to Random Forest, then XGBoost, then an Ensemble of multiple ML models. With each iteration, predictions (lead scores) became more accurate. But, I could no longer understand why the model was predicting (e.g. Unlike Linear Models, XGBoost has no coefficients).
