Global explainability aims at making the overall ML model transparent and comprehensive

Local explainability focuses on explaining the model’s individual predictions

Interpretability has to do with how accurately a machine learning model can associate a cause to an effect whereas explainability has to do with explaining the ability of the parameters hidden in deep neural nets to justify the results

Feature importance is a technique that explains the features that make up the training data using a score (importance) - to indicate how useful or valuable the feature is relative to other features. it provides the following benefits:

• variable selection
• target / label or data leakage in your model

Vertex Explainable AI

Feature attribution - indicate how much each feature in your model contributed to the predictions for each given instance

• Sampled Shapley - provides a sampling approximation of exact Shapley values to determine the black box feature importance

  

• Integrated gradients

  • mostly used in deep neural networks with image use cases (i.e. featureful images)
    • the gradient calculated will inform whcih pixel has the strongest effect on the model’s predicted class probabilities
    • provide local feature importance

• XRAI (explanation with ranked area integrals)

  

Differentiable models - you can calculate the derivative of all the operations in your TensorFlow graph. Use the integrated gradients method

Nondifferentiable models - such as operations that perform decoding and rounding tasks. Use the sampled Shapley method

Detecting bias and fairness in data

• Explainable AI feature attributions technique (present in AutoML tables to detect bias and fairness in tabular or structure dataset)
• Vertex AI to inspect models through What-If tool using the features overview functionality
• Language Interpretability Tool (LIT) to detect bias and fairness in NLP models