Repository navigation
Home
Welcome to the official technical documentation and engineering reference for ENIAD Explainable AI (XAI) Suite.
Engineering Suite for Machine Learning Model Interpretability: PDP, ALE, KernelSHAP, TreeSHAP, LIME and Grad-CAM Saliency Maps.
Developed within the State Engineering Degree in Artificial Intelligence & Digital Systems at the Γcole Nationale d'Intelligence Artificielle et du Digital (ENIAD), Berkane, Morocco.
| Chapter | Description | Primary Topics |
|---|---|---|
| Architecture | Deep architectural design & component topology | Mermaid diagrams, component interactions, runtime environment |
| Getting-Started | Developer setup & workstation configuration | Prerequisites, toolchain setup, execution commands |
| Curriculum-Guide | Detailed syllabus and laboratory breakdown | Lab objectives, expected outputs, deliverables |
graph TD
A[Raw Dataset / Input Features] --> B[Black-Box ML Model]
B --> C{Explainability Engine}
C -->|Global Interpretability| D[PDP & ALE Plots]
C -->|Local Attribution| E[TreeSHAP & KernelSHAP]
C -->|Perturbation Analysis| F[LIME Local Surrogate]
C -->|Computer Vision| G[Grad-CAM Saliency Maps]
D --> H[Auditing & Fairness Diagnostics]
E --> H
F --> H
G --> H
style A fill:#00D9FF,stroke:#333,stroke-width:1px,color:#000
style B fill:#FF6B00,stroke:#333,stroke-width:1px,color:#fff
style C fill:#3C873A,stroke:#333,stroke-width:1px,color:#fff
style H fill:#7928CA,stroke:#333,stroke-width:1px,color:#fff
Partial dependence plots show the marginal effect one or two features have on the predicted outcome of a machine learning model. A partial dependence plot can show whether the relationship between the target and a feature is linear, monotonic or more complex.
ALE plots describe how features influence the prediction of a machine learning model on average. ALE plots are faster and unbiased alternatives to partial dependence plots (PDPs) when features are correlated.
SHAP explains the prediction of an instance
LIME explains individual predictions by approximating black-box models locally with an interpretable surrogate model (such as a sparse linear regression or decision tree).
Maintained with β€οΈ by Oussama EL HADJI β’ ENIAD Berkane