Prediction API overview

The Prediction API is another machine learning APIs that's embedded in Business Central. It first and foremost supports the Late Payment Prediction extension. However, it can be used to solve other problems with regression or classification, like customer churn, quote conversion, or shipment delay.

Prediction Model

The Prediction Model for Business Central lets you easily train, evaluate, visualize models for prediction purpose. This model uses R code to perform all the tasks. The structure of the input dataset is flexible and can accept up to 25 features.

Input Data Schema

When you call the API, you need to pass several parameters:

  • method (String) – Required parameter. Specifies the Machine Learning procedure to be used. The model supports the following methods:
    • train (system decides whether to use classification or regression based on your dataset)
    • trainclassification
    • trainregression
    • predict
    • evaluate
    • plotmodel

Based on selected method you might need extra parameters, such as

  • The train_percent (Numeric) – Required for train, trainclassification, and trainregression methods. Specifies how to divide a dataset into training and validation sets. 80 means 80% of dataset are used for training and 20% for validation of result.
  • The model (String;base64) - Required for predict, evaluate, and plotmodel methods. Contents serialized model, encoded with Base64. You can get model as result of run train, trainclassification, or trainregression methods.
  • The captions (String) – Optional parameter used with the plotmodel method. Contains comma separated captions for features. If not passed Feature1..Feature25 are used.
  • The labels (String) – Optional parameter used with the plotmodel method. Contains comma separated alternative captions for labels. If not passed actual values are used.
  • The dataset - Required for train, trainclassification, trainregression, evaluate, predict, and it consists of:
    • Feature1..25 – The features are the descriptive attributes (also known as dimensions) that describe the single observation (record in dataset). It can be integer, decimal, Boolean, option, code, or string.
    • Label – Required but should be empty for predict method. The label is what you're attempting to predict or forecast

Output Data Schema

The output of the service:

  • The model (String;base64) – Result of execution of train, trainclassification, and trainregression methods. Contains serialized model, encoded with Base64.
  • The quality (Numeric) – Result of execution of train, trainclassification, trainregression, and evaluate methods. In current experiment we use Balanced Accuracy score as a measure of a model’s quality.
  • The plot (application/pdf;base64) – Result of execution of plotmodel method. Contains visualization of model in pdf format encoded with Base64.
  • The Dataset – Result of execution of predict method, consists of:
    • Feature1..25 – same as input.
    • Label – the predicted value
    • Confidence – the probability that classification is correct.

Prediction API

All logic of the Prediction API is concentrated in the ML Prediction Management codeunit (ID=2003) and has the following methods:

For Business Central online, the experiment is published by Microsoft and connected to the Microsoft subscription. For other deployment options, you have to create Machine Learning resources in your own Azure subscription. You can find sample steps in the sample repo.

The purpose of this task is to get the API URI and API key and pass them into the Initialize method. That gives the Prediction API the end-point to contact:


var
  MLPredictionManagement: Codeunit "ML Prediction Management";
  URITxt: TextConst ENU = 'https://../execute?api-version=2.0&details=true';
  KeyTxt: TextConst ENU = 'TlfjUL..Mnrahg==';
begin
  MLPredictionManagement.Initialize(URITxt, KeyTxt, 0);

In Business Central online, you can use the default credentials. In that case, you can use the following method instead:

  MLPredictionManagement.InitializeWithKeyVaultCredentials(0);

Note

You can always switch back to resources managed by Microsoft by removing values from API URL and API Key fields.

Once initialized, you must prepare the training dataset. Just like the Forecasting API, the Prediction API can take almost any record as input. But, from a practical perspective, it's recommended to create a buffer table to aggregate the training data. In this case, you can gather data from multiple sources and perform the data transformation as needed. Even in the simple “Christmas Apparel Demo”, the data was coming from multiple sources: Sales prices come from the Item card or the Sales Prices table; gender, material, and sleeve length came from item attributes. So let’s get started by creating the buffer table.


Table 50136 "ML Prediction Parameters"
{
    fields
    {
        field(1; Counter; Integer) { AutoIncrement = true; }
        field(2; Price; Option) { OptionMembers = Low,Medium,High; }
        field(3; Gender; Option) { OptionMembers = Man,Women; }
        field(4; Material; Option) { OptionMembers = Cashmere,Silk,Wool,Acrylic,Viscose,Cotton; }
        field(5; SleeveLength; Option) { OptionMembers = Full,Half,Short,Threequarter,Butterfly,Sleveless; }
        field(10; DecemberSales; Option) { OptionMembers = Low,Medium,High; }
        field(11; Confidence; Decimal) { }
    }

    Keys
    {
        key(PK; Counter) { Clustered = true; }
    }
}

It’s just a buffer table, where Counter is the primary key. Price, Gender, Material, and SleeveLength are features. The DecemberSales field stores labels for training and receives the predicted value. Confidence is a field that specifies the probability that the classification is correct.

The next step is to fill the buffer table with data, done in AL and not shown here.

Price Gender Material SleeveLength DecemberSales
Medium Woman Cotton Threequarter Low
High Man Cotton Half Low
Medium Woman Acrylic Butterfly High
Medium Man Silk Short Medium
Medium Woman Cotton Butterfly High
Low Woman Acrylic Threequarter Low
Medium Man Wool Full Medium
Low Man Cotton Full High
Low Woman Acrylic Short High

Once we have the training data, we can use the Prediction API to train the model. Let see if the system figures out how different combinations of price, gender, material lead to different sales volumes in December. We use the SetRecord method to point to the table with the training dataset. After that, we run the AddFeature method for each field that contains our features (Price, Gender, and so on). We can add up to 25 features.

Then, we must specify, which fields contain the label (the expected output) and confidence. Though confidence isn't needed at this stage, the way the API is implemented requires us to specify the confidence field.

var
  MLPredictionParameters: Record "ML Prediction Parameters";
  ModelAsText: Text;
  ModelQuality: Decimal;
begin
  MLPredictionManagement.SetRecord(MLPredictionParameters);

  MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Price));
  MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Gender));
  MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Material));
  MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(SleeveLength));

  MLPredictionManagement.SetLabel(MLPredictionParameters.FieldNo(DecemberSales));  

  MLPredictionManagement.SetConfidence(MLPredictionParameters.FieldNo(Confidence));

The last method to call is Train, which sends data to the Azure Machine Learning experiment and receives the trained model as text and an indication of the quality of model.

  MLPredictionManagement.Train(ModelAsText, ModelQuality);

If the quality of the model (the percentage of correct predictions) is acceptable for your business scenario (0.8 or higher), then you can store the model for future use, for example in a BLOB field. Now, you have the trained model, and you can use it for classification tasks.

The application code is similar to the code we wrote for training purposes. Again we use a buffer table, but this time it contains records with features only. Notice that in this case, you keep the label field empty (in our scenario, that’s the DecemberSales field).

MLPredictionManagement.Initialize(URITxt, KeyTxt, 0);

MLPredictionManagement.SetRecord(MLPredictionParameters);
MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Price));
MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Gender));
MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(Material));
MLPredictionManagement.AddFeature(MLPredictionParameters.FieldNo(SleeveLength));
MLPredictionManagement.SetConfidence(MLPredictionParameters.FieldNo(Confidence));
MLPredictionManagement.SetLabel(MLPredictionParameters.FieldNo(DecemberSales));

MLPredictionManagement.Predict(ModelAsText);

The last method that you call is the Predict method, which sends the model and data to the Azure Machine Learning experiment. As soon as results are received, the fields DecemberSales and Confidence are updated with the predicted class and the probability that the classification is correct.

Now, you can loop through the updated buffer table and read the label and confidence for each record used in the prediction. For more information, see the source code of the Late Payment Prediction extension.

See also

Forecasting API overview
The Late Payment Prediction Extension