Explainable AI

Research Papers, Blogs & Resources

1. Explaining the Predictions of Any Classifier
2. Integrated Gradients
3. Robustness of Interpretability Methods
4. Interpretable Machine Learning Web Book
5. LIME TDS 1 | LIME Blog | LIME Text Explain

Python Libraries

1. explainerdashboard
2. omnixai
3. InterpretML
4. ELI5
5. Shapash
6. LIME
7. SHAP

Local vs Global Interpretability

• Global feature importance comes from understanding how the model breaks down the global space/dimension and tries to explain how are all the points in data predicted
• Local feature importance comes from understanding how the model behaves around the locality of current query point
• Within the neighbourhood of the query point create a surrogate model which aproximates the original model in that neighbourhood. This surrogate model can be a basic linear model or a Decision Trees, which have easily interpretable feature importance
• Interpretable representation : Conversion of a feature of dimension d to a binary feature of dimension d'. Ex : for text use Bag of Word, for image use super pixel, for real values use real valued binning
• The Surrogate model will use these Interpretable representation as input rather than the original input