Summary

In this chapter, we explored a powerful XAI tool. We saw how to analyze the features of our training and testing datasets before running an ML model.

We saw that Facets Overview could detect features that bring the accuracy of our model down because of missing data and too many records containing zeros. You can then correct the datasets and rerun Facets Overview.

In this iterative process, Facets Overview might confirm that you have no missing data but that the data distributions of one or more features have high levels of non-uniformity. You might want to go back and investigate the values of these features in your datasets. You can then either improve them or replace them with more stable features.

Once again, you can rerun Facets Overview and check the distribution distance between your training and testing datasets. If the Kullback-Leibler pergence is too significant, for example, you know that your ML model will produce many errors.

After several iterations and a lot of fine-tuning, Facets Overview provides the XAI required to move on and use Facets Dive.

We saw that Facets Dive's interactive interface displays the data points in many different ways. We can choose the way to organize the binning of the x axis and y axis, providing critical insights. You can visualize the data points from many perspectives to explain how the labels of your datasets fit the goals you have set.

In some cases, we saw that the counterfactual function of Facets Dive takes us directly to data points that contradict our expectations. You can analyze these discrepancies and fine-tune your model or your data.

In the next chapter, Microsoft Azure Machine Learning Model Interpretability with SHAP, we will use the XAI Shapley value algorithms to analyze ML models and visualize the explanations.