AutoML real-time predictions

This step-by-step guide helps you migrate your application from the deprecated AutoML real-time predictions to the new real-time prediction endpoint in the Machine Learning API.

Using the Machine Learning API, you can now add multiple models to a deployment. A system of aliases is used in deployments to allow for dynamic swapping of models for use in predictions. For more information, see Using multiple models in your ML deployment on Qlik Help.

Understand the endpoint changes

The new real-time prediction endpoint in the Machine Learning API replaces the deprecated AutoML real-time predictions endpoint.

The main change is the response format: responses are now wrapped under data.attributes. The request body format is unchanged.

Here’s how your requests will change:

Example response from the deprecated endpoint:

{
  "schema": [
    { "name": "feature_1", "type": "numeric" },
    { "name": "feature_4_predicted", "type": "categorical" },
    { "name": "feature_4_no", "type": "numeric" },
    { "name": "feature_4_yes", "type": "numeric" },
    { "name": "not_predicted_reason", "type": "categorical" }
  ],
  "rows": [
    [0, "yes", 0.50, 0.49, null],
    [1, "no", 0.76, 0.23, null]
  ]
}

Example response from the new endpoint:

{
  "data": {
    "type": "realtime-prediction",
    "attributes": {
      "schema": [
        { "name": "feature_1", "type": "numeric" },
        { "name": "feature_4_predicted", "type": "categorical" },
        { "name": "feature_4_no", "type": "numeric" },
        { "name": "feature_4_yes", "type": "numeric" },
        { "name": "not_predicted_reason", "type": "categorical" }
      ],
      "rows": [
        [0, "yes", 0.50, 0.49, null],
        [1, "no", 0.76, 0.23, null]
      ]
    }
  }
}

Prerequisites

Before you begin, make sure you have:

• API credentials with permission to access the Machine Learning API endpoints. For more information, see Authentication.
  Required permissions and scopes

  If you use an API key: The user or service account must be assigned the Automl Deployment Contributor role to call the real-time prediction endpoint.
  If you use OAuth: Your token must include the automl-deployments:predict scope, which enables the required permissions for the real-time prediction endpoint.

• A deployed and activated model for making predictions.
  • Your deployment must be created with "enablePredictions": true to allow real-time predictions.
  • Your model must be activated on the deployment using POST /ml/deployments/{deploymentId}/actions/activate-models.
• The deploymentId: The unique identifier for your deployment.

Make a real-time prediction request

Once your model is activated, and you have the required input schema, you can make predictions using the new real-time prediction endpoint in the Machine Learning API.

Send a real-time prediction request with a request body that includes the schema and rows that match the features expected by your model.

curl -X POST "https://<TENANT_URL>/api/v1/ml/deployments/{deploymentId}/realtime-predictions/actions/run" \
  -H "Authorization: Bearer {TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "schema": [
      {"name": "feature_1"},
      {"name": "feature_2"},
      {"name": "feature_3"},
      {"name": "feature_4"}
    ],
    "rows": [
      [0, "France", 5, "yes"],
      [1, "Germany", 20, "no"]
    ]
  }'

(Optional) Add query parameters:

• includeShap=true to include SHAP values in the response (explainability).
• includeSource=true to include source data in the response.
• includeNotPredictedReason=true to include reasons for rows that could not be predicted.
• index to specify the name of the feature in the source data to use as an index in the response data.

curl -X POST "https://<TENANT_URL>/api/v1/ml/deployments/{deploymentId}/realtime-predictions/actions/run?includeShap=true" \
  -H "Authorization: Bearer {TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "schema": [
      {"name": "feature_1"},
      {"name": "feature_2"},
      {"name": "feature_3"},
      {"name": "feature_4"}
    ],
    "rows": [
      [0, "France", 5, "yes"],
      [1, "Germany", 20, "no"]
    ]
  }'

Verify the response

If your request is successful, you’ll receive a JSON response like this:

{
  "data": {
    "type": "realtime-prediction",
    "attributes": {
      "schema": [
        { "name": "feature_1", "type": "numeric" },
        { "name": "feature_4_predicted", "type": "categorical" },
        { "name": "feature_4_no", "type": "numeric" },
        { "name": "feature_4_yes", "type": "numeric" },
        { "name": "not_predicted_reason", "type": "categorical" }
      ],
      "rows": [
        [0, "yes", 0.50, 0.49, null],
        [1, "no", 0.76, 0.23, null]
      ]
    }
  }
}

The response format differs between the deprecated and new endpoints:

• Deprecated endpoint: The response contains schema and rows at the root level.
• New endpoint: The response contains schema and rows inside the data.attributes object.

Troubleshooting

If you get an error, see the following common issues:

• 401 Unauthorized: Verify your API credentials and permissions.
• 404 Not Found: Make sure your deploymentId and modelId are correct, and your model is deployed and activated.
• 400 Bad Request: Verify your request body for correct schema and data types.

For more information, see the Machine Learning API reference.