After the algorithm model is trained, you can use the algorithm model to predict repurchases and obtain the purchase probability of users N days in the future.
"N days in the future" refers to the latest behavior time of the behavior dataset used by the algorithm model. The value of N has been set when the algorithm model is created. For example, if today is 20210910, the behavior time range of the behavior dataset is 20190101 to 20210801, and the algorithm model sets N to 60, then the prediction will be made for 60 days from 20210802.
Create a prediction task
When you create a prediction task, a unique algorithm model in the Trained state is used. Make sure that an algorithm model in the Trained state is available in this workspace. Otherwise, you cannot create a prediction task.
The range of predictable users is the users involved in the behavioral dataset used by the algorithm model, and the users must have purchased the behavioral dataset within the past year.
The users that participate in the prediction must be saved as an Audience. The ID of the audience must be of the same type as the user ID in the behavior dataset. If the audience includes users outside the scope of the preceding behavioral dataset, these users will not participate in the prediction.
The number of used prediction tasks /number of purchased prediction tasks is displayed at the top of the list, which is the sum of all spaces in the organization. New tasks consume the number of available predicted tasks. Failed executions are not counted.
Create a prediction task
When you create a prediction task, a unique algorithm model in the Trained state is used. Make sure that an algorithm model in the Trained state is available in this workspace. Otherwise, you cannot create a prediction task.
The range of predictable users is the users involved in the behavioral dataset used by the algorithm model, and the users must have purchased the behavioral dataset within the past year.
The users that participate in the prediction must be saved as an Audience. The ID of the audience must be of the same type as the user ID in the behavior dataset. If the audience includes users outside the scope of the preceding behavioral dataset, these users will not participate in the prediction.
The number of used prediction tasks /number of purchased prediction tasks is displayed at the top of the list, which is the sum of all spaces in the organization. New tasks consume the number of available predicted tasks. Failed executions are not counted.
Procedure
Choose Workspace> User Insight> Model Center> Repurchase Prediction> Prediction Task.
Click New Task in the upper-right corner. The following figure shows the configuration page.
Enter a task name and select Audience as the source of the people who participate in the prediction.
Select Confirm that the new task will consume available prediction tasks and click OK. The system starts the prediction task.
Manage prediction tasks
The following figure shows the task list.
The task execution status is as follows:
Pending: When the number of model training and crowd prediction tasks in progress in the organization exceeds five, the excess tasks are queued for execution.
The command is being run.
The command is run.
Execution failure: If no results are returned for 24 hours, the execution is automatically stopped, or if the source audience does not contain any predictable users, the execution fails. Moving the mouse over
the icon displays the reason for the failure. Unpublished: The task is automatically unpublished 180 days after it is created. The task data is deleted but the list records are retained.
NoteWhen you undeploy a task, if the corresponding model has been unpublished and no other tasks are associated with the model, the model will be deleted at the same time.
You can view the result details and remove tasks.
View the result details
If the task is successfully executed, you can click the 
icon to view the prediction results of the task. For more information, see View Result Details.
Remove
For a task that is not unpublished, you can click the 
icon to delete the task.
When you delete a task, if the corresponding model has been unpublished and the model has no other associated tasks, the model will be deleted at the same time.
View the result details
For a successful task, click Result Details to go to the details page and view the task information and detailed results.
The information about the data migration task.
The task information is displayed at the top of the details page, as shown in the following figure.
Pay attention to the Actual Number of Users field. This field indicates the number of users in the audience who participated in the prediction. The rest of the users in the audience failed to participate in the prediction and are invalid users. The requirement for participating in the prediction is that the user ID is matched in the behavior data set used by the algorithm model, and the user has a purchase behavior record in the past year.
When you filter audiences from the prediction results, you must refer to the Actual Number of Predicted Audiences parameter. For more information, see Detail Result.
Detail Results
The Details tab displays the following results:
Example purchase probability for a user in the next N days:
The score distribution of purchase probability in the next N days. Move the pointer over the score graph. The pop-up window displays the number of users who have obtained the current score and the total number of users who have obtained the current score greater than or equal to the current score.
Click New Audience to save the predicted users as an audience.
Determine which users need to be saved as audiences:
Save the users with the highest purchase probability of the specified number as the audience, as shown in the following figure.
We recommend that you set a small number of people (for example, the actual number of people × 5% or 10%) to improve the prediction effect. If you need a larger number of people, we recommend that you refer to the Model Verification results and select the N % that corresponds to the accuracy and recall rate of a high potential verification group. Then, you can obtain the number of people by actual number of people × N %.
Conversely, if you have determined how many people need to be saved, you can use the number of people to extrapolate the expected prediction accuracy and recall rate.
Save users with the specified purchase probability interval as an audience, as shown in the following figure.
Select the ID type of the audience to be saved. You can save the user ID of the behavioral dataset used by the algorithm model. You can also save other ID types of the crowd source audience for the behavioral dataset or prediction task. Select an ID type that is convenient for subsequent marketing. If a user does not have this ID type data, it cannot be saved to the generated audience.
Enter the name and description of the audience, select the folder where the audience is saved, and select a child campaign. For more information, see Marketing Campaign Documentation.
Click OK to create the audience, or click OK to go to the Audience Management page.
On the Audience Management page, the Repurchases prediction field displays the source of the audience. You can analyze, push, and download the content. For more information, see Audience Basic Functions and Audience Push Functions. You can also create marketing tasks for audiences. For more information, see User Marketing Documentation and Automatic Marketing Documentation.
