Integrating a model in your app

Sample data structure

In this example, the Entity is SampleCustomerData with the following Attributes:

• CustomerId
• CreditScore
• Geography
• Gender
• Age
• Tenure
• Balance
• NumOfProducts
• HasCrCard
• IsActiveMember
• EstimatedSalary
• Churned

You can train a model for the attribute prediction, and predict a binary decision (True or False) for Churned. The score in this example returns the likelihood on the scale from 0 to 1, or 0% to 100%.

You should exclude CustomerId (from your customer data) and Id (Entity record identifier) as they don't contribute to the model. Check also how to prevent data bias.

Sample logic

After you reference the plugin, create logic that asks your model for predictions. In this example, there's GetAttributePrediction that uses a record of an Entity as the input.

1. Prepare the data for sending it to the model. As the model accepts structured JSON data, use *JSONSerialize to convert a record from the Entity to a JSON.

   

   In the example, the action requires an input parameter of the data type SampleCustomerData, which is then the Data for JSONSerializeCustomer.

2. Send the data to the model by using Logic > Service Actions > MLBuilderPlugin > ModelScore > GetAttributePrediction action with JSONSerializeCustomer.JSON as the ScoreInputs value.

   

   You also need to enter the name of the model in the property ModelName. Get the name of the model you want to use from the list of trained and deployed models in ML Builder.

   Notes

   • Make sure you deploy a model you want to use in your app.
   • The ModelName in the actions and the model name in ML Builder must be the same. For example, if the model name in ML Builder is My Model, then enter My Model in the ModelName field.

3. Check if the model returns data. The value of GetAttributePrediction.IsSuccess should be True. In the example, the shows an error if the request to the model isn't successful.

   

4. Get the top result of the prediction. Do that by evaluating GetAttributePrediction.TopPrediction[0].score and assigning the value to a variable. In the example, the variable is TopPrediction.

   

   The data type of TopPrediction is decimal because you're predicting the likelihood of the customer ending the contract (churning), on a scale from 0 to 1. This is because you trained the model with the data where the churn value is 0 or 1.

5. Decide whether you want to show the prediction by checking how "confident" the prediction is. For example, you can show results when the top score is above 0.3, or 30%.

   

6. Finally, update the user interface to show the result. In a production app you could show customer data in red if there's a risk of churn, or green, if there's no risk. For testing, you can use notification with the prediction result.

Steps overview

Here is the summary of the logic. The numbers match the steps of the instructions.

Service actions

The available service actions are in Logic > Service Actions > MLBuilderPlugin > ModelScore. Here is more information about the actions.

Parameters

Here is more information about the parameters you can use in the ML Builder plugin actions.

Structures

You can check the structures in Data > Structures > MLBuilderPlugin.

The meaning of class and score in the ScoreItem depends on the prediction:

Here are some examples:

When you want to predict a numeric value, the class in the response is always prediction:

{"class": "prediction", "score": 12.567}

If you predict a boolean or a value from a set of values, class in the response for ScoreResult holds a value, while score indicates probability:

{"class": "Green", "score": 0.25}

{"class": "Yellow", "score": 0.35}

{"class": "Blue", "score": 0.40}

{"class": "True", "score": 0.40}

{"class": "False", "score": 0.60}

In TopPrediction, class is a boolean or a value from a set of values, and the score is the probability.

{"class": "Blue", "score": 0.40}

{"class": "False", "score": 0.60}