Types of Explainability in AI Systems: Post-hoc and Intrinsic

Intermediate

🧩 Categories of Explainability in AI

Explainability in AI can be divided into two main categories:

🧠 1. Post-Hoc Explainability

Generates explanations after a model makes a decision.
Applicable to complex, black-box models.

🔧 Techniques:

  • 📊 Feature importance
  • 🧪 Surrogate models
  • 🧬 Example-based explanations

🧰 Common Tools:

  • 🧮 SHAP (SHapley Additive Explanations)
  • 🔍 LIME (Local Interpretable Model-agnostic Explanations)

📌 Example:
Explaining a deep neural network using SHAP or LIME.

🟢 2. Intrinsic Interpretability

Models that are inherently understandable and transparent by design.

🧠 Examples:

  • 🌳 Decision Trees
  • ➕ Linear Regression / Logistic Regression

📌 Example:
A decision tree clearly shows the decision path for each prediction.

🎯 Choosing the Right Approach

Understanding these categories helps in selecting suitable techniques based on:

  • ⚙️ Application complexity
  • 📈 Performance needs
  • 🔍 Transparency requirements