Work with predictions
Dynamics 365 Customer Insights - Data comes with various options that leverage AI and machine learning to predict data.
Predictions offer capabilities to create better customer experiences, improve business capabilities, and revenue streams. We strongly recommend you balance the value of your prediction against the impact it has and biases that might be introduced in an ethical manner.
Predictions and custom models that have been defined for your organization can be accessed from the Intelligence area of the application.
Out-of-the-Box, Customer Insights – Data includes four prebuilt models and two options for creating your own custom models.
Predictions
The easiest way to start with predicting data is to leverage predefined models. They only require certain data and structure to generate insights quickly.
The following models are available:
Customer lifetime value: Predicts the potential revenue of a customer throughout the entire interaction with a business. See: Predict Customer lifetime value.
Product recommendation: Suggests sets of predictive product recommendations based on purchase behavior and customers with similar purchase patterns. See: Predict product recommendations.
Subscription churn: Predicts whether a customer is at risk for no longer using your company’s subscription products or services. See: Predict subscription churn.
Customer Sentiment analysis: Analyze sentiment of customer feedback and identify business aspects that are frequently mentioned. See: Analyze sentiment in customer feedback.
Define predictions
Customer Insights - Data allows you to create predictions in two different ways:
From the Customer entity
From Quick Create segments
Regardless of the method that you use to create the prediction, they're always based on a field that you want to make the prediction on. For example, a customer's gender might have only been included in one of your data sources. Because of this, you might have multiple customers who are missing gender information in their profile. You could create a prediction to determine whether a customer is male or female.
You can create a prediction on the customer profile by going to Data and selecting Entities. Open the Customer entity under Profiles. Under the Summary column, locate the attribute name that you want to predict values for and then select the Overview icon.
Customer Insights - Data will display the overview window for the field that you selected. This window will provide additional information that is related to the field, such as the number of unique and total records and if any records are missing this data. If a high rate of missing values is shown for your attribute, you can select Predict missing values to make a prediction.
One reason that Dataverse is required to make a prediction is because the prediction results are stored in a custom Dataverse entity that is automatically created. It contains the results of the prediction and the ID of the record that the result is for. You'll need to provide a prediction name (displayed in Customer Insights - Data) and an Output entity name (name of the Dataverse entity to be created) for the results of the prediction.
A prepopulated list of options will show based on the field that you selected. Depending on what you're trying to map, the field options will vary. The following image shows that the only category options are 0 or 1 because they map to the true/false or binary nature of the prediction.
In the Category column, map the field values that you want to be classified as "0" in the final prediction to 0, and map the items that you want to be classified as "1" in the final prediction to 1.
After you've created your model, it will start processing, which can take some time depending on the size and complexity of data. The results will be available in the new entity that was created based on the Output entity name of the prediction that you created.
Predictions that are created in Customer Insights - Data can be modified to adjust the effectiveness of the model. For example, the model might be looking at fields that aren't the best fields for making a prediction based on the field that you're working with. Removing those fields from the model might provide more accurate results. Predict models that were created in Customer Insights - Data can be opened and edited in AI Builder by viewing the prediction and selecting Customize in AI Builder.
Note
Only models that were created by using the Predict option in Customer Insights - Data will display in the Customer Insights - Data user interface. Any new models that are created directly in AI Builder are not displayed.
Create a prediction while creating a segment
Another way that you can create predictions for missing values on a specific attribute is when you're creating a quick segment that is based on either your unified Customer entity or Customer_Measure entity. You can complete this process by going to Segments, creating a new segment, and then selecting Create from Profile.
You'll create the segment as you would any other quick segment. You'll need to define a field to create a segment on and then select the operator that you want to use. If the segment that you created has incomplete data in the source field, the system will ask if you want to predict the missing values.
View predictions
Regardless of the method that you used to create a prediction, after it has been created and processed, you can view the results by going to Predictions under the Intelligence heading. Any configured predictions will be available under the My Predictions tab. When you open a prediction, several visuals will be available to help you understand it better.
Predicted values - Shows mapping that was created during the Field to Category mapping phase.
Top influencers - Factors within your dataset that were most likely to influence the prediction's confidence of your Field value being mapped to a specific category.
Performance - Indicates how the predictions are doing.
Preview - Shows samples of the output dataset from your prediction and the likelihood, or your confidence, of the predicted value where 0 is uncertain and 1 is certain.
Predictions update like other items, such as segments and measures, when scheduled refreshes occur. They can be updated in between scheduled refreshes by selecting the update icon.
Manage predictions
As your predictions run, you might encounter scenarios where the Predictions fail. There are several types of errors that can occur, and they describe what condition caused the error. For example, an error that there's not enough data to accurately predict is typically resolved by loading more data into Customer Insights. In these scenarios, you can troubleshoot your predictions from the My predictions tab and select log.
One way that you can gain insights related to errors and warnings generated by your out-of-box predictions is through the Input data usability report. The report is available after a model has completed its training process. It's created for each model separately, regardless of if it completed successfully or not. It also gives recommendations on how to improve the model performance.
After an out-of-box model has completed its training step, view the report:
Select the ellipses next to the model's name and choose Input data usability report.
Select the status of a model select See Input data usability report in the side pane.
Selecting one of the models in the list and select Input data usability report in the command bar.
Open the model results page and select Input data usability report in the command bar.
The following columns in the report contain helpful information to improve the data for the model.
Name: Descriptive name of the error, warning, or recommendation.
Step: Model phase, train, or score, the information refers to.
State: Severity of the information (error, warning, recommendation).
Column name: Column in an entity that needs to be modified to improve the model performance.
Entity name: Name of the entity that needs to be modified to improve the model performance.
Details: Details about the error, warning, or recommendation.
For more information, see Complete your partial data with predictions.