Edit

Create a prediction model

  • Article

This example creates a Power Apps prediction AI model that uses the Online Shopper Intention table in Microsoft Dataverse. To get this sample data into your Microsoft Power Platform environment, enable the Deploy sample apps and data setting when you create an environment as described in Build a model in AI Builder. Or, follow the more detailed instructions in Data preparation. After your sample data is in Dataverse, follow these steps to create your model.

  1. Sign in to Power Apps or Power Automate.

  2. On the left pane, select ... More > AI hub.

  3. Under Discover an AI capability, select AI models.

    (Optional) To keep AI models permanently on the menu for easy access, select the pin icon.

  4. Select Prediction - Predict future outcomes from historical data.

  5. Select Create custom model.

Select your historical outcome

Think of the prediction you want AI Builder to make. For example, for the question "Will this customer churn?", think about questions like these:

  • Where is the table that contains information about customer churn?
  • Is there a column there that specifically states whether the customer has churned?
  • Are there unknowns in a column that might cause uncertainty?

Use this information to make your selections. Working with provided sample data, the question is "did this user who interacted with my online store make a purchase?" If they did, there should be revenue for that customer. Therefore, whether there's revenue for this customer should be the historical outcome. Wherever this information is empty is where AI Builder can help you make a prediction.

  1. In the Table dropdown menu, select the table that contains the data and the outcome you want to predict. For the sample data, select Online shopper intention.

  2. In the Column dropdown menu, select the column that contains the outcome. For the sample data, select Revenue (Label). Or, if you want to try out predicting a number, select ExitRates.

  3. If you selected an option set that contains two or more outcomes, consider mapping it to "Yes" or "No" because you want to predict whether something will happen.

  4. If you want to predict multiple outcomes, use the Brazilian e-commerce dataset in the sample, and select BC Order in the Table dropdown menu and Delivery Timelines in the Column dropdown menu.

Note

AI Builder supports these data types for the outcome column:

  • Yes/No
  • Choices
  • Whole number
  • Decimal number
  • Floating point number
  • Currency

Select the data columns to train your model

After you select the Table and Column and map your outcome, you can make changes to the data columns used to train the model. By default, all relevant columns are selected. You can deselect columns that might contribute to a less accurate model. If you don't know what to do here, don't worry. AI Builder will try to find columns that provide the best model possible. For the sample data, just leave everything as is and select Next.

Data column selection considerations

The most important thing to consider here is whether a column that isn't your historical outcome column is indirectly determined by the outcome.

Let's say you want to predict whether a shipment is going to be delayed. You might have the actual delivered date in your data. That date is only present after the order is delivered. So, if you include this column, the model will have close to 100 percent accuracy. The orders that you want to predict won't have been delivered yet, and won't have the delivered date column populated. So, you should deselect columns like this before training. In machine learning, this is called target leakage or data leakage. AI Builder tries to filter columns that are "too good to be true," but you should still check them.

Note

When you're selecting data fields, some data types—like Image, which can't be used as input to train the model—aren't shown. In addition, system columns like Created On are excluded by default.

If you have related tables that might improve the performance of the prediction, you can include those as well. As you did when you wanted to predict whether a customer will churn, you should include additional information that might be in a separate table. AI Builder supports many-to-one relationships at this time.

Filter your data

After you select data columns for training, you can filter on your data. Your tables will contain all rows. However, you might want to concentrate on training and predicting on a subset of rows. If you know that there's irrelevant data within the same table you're using to train a model, you can use this step to filter it.

For example, if you apply a filter to look at only the US region, the model will train on rows where the outcome is known only for the US region. When this model is trained, it will only make a prediction for rows where the outcome is not known for only the US region.

The filtering experience is the same as in the Power Apps view editor. Start by adding:

  • A row, which contains a single filter condition.
  • A group, which allows you to nest your filter conditions.
  • A related table, which allows you to create a filter condition on a related table.

Select the column, the operator, and the value that represents a filter condition. You can use the checkboxes to group rows or to bulk-delete rows.

Next step

Train and publish your prediction model