Actionable-Insights

Actionable Insights can be found in 2 different locations:

1) In the model view under Features section: For a particular model, the default conditions for different features such as Lower bound, Upper bound, Cost of Change etc. can be configured in this section.

2)  In the model view under Deployment section: This takes us to Actionable insights demo page where feature values can be updated and the corresponding change in the prediction can be checked.

The usual approach is to first set the default feature configurations at the model level and then change additional requirements for the features in the demo page, based on the individual case. 

Actionable Insight Under Feature Section:

In this category, a screen is displayed with all the features of the selected model. "Actionable Insight " configuration and the membership function chart is shown for the selected feature. This page allows us to edit the configuration of each feature according to the requirement. In the models features section, each feature has actionable insights configurations.

1. The first is whether the feature is a design variable.  This means this feature will be modified in the actionable insights to produce the desired outcome.
2. The second is either the upper and lower bounds of the continuous feature or the categories of the categorical feature
3.  The third setting is the cost of change, (which can be "Low" "Medium" or "High") which is a selection to capture the cost of changing this feature. For example, the cost of changing the loan amount might be "low" because the user might be willing to accept a different loan amount in their application. However the cost of moving to a different state might be "high". Meaning its not something the user wants to do lightly, and may only action this suggestion if there is no other way of getting the desired outcome.
4. Finally. You must save the changes for them to take affect.

In case of continuous features, lower bound and upper bound needs to be specified whereas in categorical features, category values box to which values can be changed needs to be checked.

Actionable Insights Under Deployment Section:

This takes us to the demo page of Actionable insights of the model. This page is used to test the actionable insights generated by the model.

1.  We have a customer who is classified as Uncreditworthy, and they want to know what they can do to be classified as Creditworthy. Their Uncreditworthiness is 59.59% so is close enough to the 50% threshold to have a chance of reclassification.
2. From this specific customer profile, we can select the cog icon to open this customer in the Inference Page. Then from here, we can choose the Actionable Insights tab.
3. We should make sure the Data Source option is Live.
4. The target class is the classification outcome that we want for this customer. In this case we can choose between Creditworthy and Uncreditworthy.

   

5. The target value is the % classification of the Minority Class (in this case Uncreditworthy). In a perfect scenario, we would aim for a target value of 0 (0% Uncreditworthy). However, this may be unachievable. We only need the customer to have a value of < 0.5. So we can set the target value here to 0.45 (45% Uncreditworthy)

   

6. We now have the option of further configuring the features. The default values, selected and saved in the Actionable Insights Feature section, are already set here. However, for this specific customer we can further modify the features that they would be willing to change.

   

7. Once at least 1 feature is a design variable the Get Insights button becomes active. This will consume all the feature configurations for features that have been chosen as design variables.

   

8. After a short delay, a response will be returned with the suggested changes the customer needs to make, and the new classification they would receive if they made these changes.
9. By clicking Accept Solution we are navigated back to the Inference page, with the new changes in place.
10. We can then click on the Inference card (where it also informs us of the new bucket number) to take us to the drivers and explain-ability.
11. We can now analyse why the customer has been reclassified as Creditworthy.

    

     Major Drivers before Reclassification

    

     Major Drivers after Reclassification

 In this case, moving out of California and bumping their credit score up from low / medium to medium is the reason for the reclassification.