Training a Model

After you add or update a Ticket Classification model, you must train the model in order for Work Center to predict field values. It is also best practice to periodically retrain models to update them using newer data. When you train a model, SBM uses the Application Engine's Machine Learning component to analyze items in the system, create field value predictions based on existing data in those items, and then validate the predictions.

Machine Learning is an Application Engine sub-component that uses historical data in your system to make field-level predictions in Work Center, which can assist end-users with completing work item forms. Because the Machine Learning training is CPU-intensive and can impact system resources that are needed to power Work Center, it can be beneficial to use a dedicated server to perform the training instead of your main Application Engine server. Alternatively, you can use the Scheduler feature in Application Administrator to schedule Machine Learning training during off-peak hours.

For details on creating a Scheduler job that trains models automatically during off-peak hours, see Working with the Scheduler.

To train a Ticket Classification model manually:

From the Ticket Classification Models view, do one of the following:
- Select an existing model, and then click Train.
- Add a new model, save your changes, and then click Train.
- Edit an existing model, save your changes, and then click Train.
Open the Log tab to view the status of the current training session. For details, refer to Training Log.

When a model is trained, the total number of items in the data set is divided into two groups: 90% of the items are reserved for training and 10% of the items are reserved for validation. The accuracy score represents prediction success for trained items when compared to items set aside for validation.

Note that smaller data sets result in lower accuracy scores. To increase the accuracy value, include more applicable items in the model's data set.

Higher accuracy scores are an indication of success, which mean you can be confident that field value predictions will be consistent with historical items in the model's data set.