As your data changes over time, the quality of the predictions from a particular model
can degrade. With IBM Watson® Machine Learning for z/OS, you can schedule
periodic evaluations of the accuracy of your models.
Before you begin
Complete the following tasks:
Procedure
-
Sign into the WML for z/OS web user interface with
your user name and password.
-
From the sidebar, navigate to the
Model Management
page and click the
Models
tab.
-
Select the TentModelVB model and from the ACTIONS menu, select
Setup
evaluation
.
-
Select an
Evaluator
, such as binary. For Use performance metrics to
monitor this model
, select areaUnderROC
and click Notify when less
than
. Specify a value or keep the default.
-
For
Data Source
, select an existing data source and data set and specify the
minimum trigger records. If needed, create a new data source.
-
If you want to use
Time range
, name the time range column and choose a
time.
-
Click Setup to complete the evaluation setup.
-
From the ACTIONS menu, click
Deploy
to deploy or redeploy the model.
-
Go to the Deployments tab and from
the ACTIONS menu for the
TentModelVB model, select Schedule evaluation.
The Schedule
page displays the evaluator and data set that you already
selected.
If a model has multiple deployments, you can schedule all of them for evaluation. However, you cannot schedule more than one deployment of the same base model for evaluation at a time. You must remove the evaluation setup from one deployment before you can schedule another one for evaluation.
-
For Starts at, specify the earliest
time that the model can be reevaluated.
The model is not reevaluated before this time.
-
For Repeat, select the frequency for
evaluation.
The first reevaluation occurs at the next scheduled repeat time. Optionally, specify the
last time when a scheduled evaluation will occur.
-
Select the automatic retain option if the result of the scheduled evaluation is poor. You can
click the deployment name to open the deployment details and review the evaluation results.
-
Click Schedule to save the scheduled evaluation.
Results
When an evaluation result for a model is below the threshold that you specified for the
evaluation, a row is shown for the model in Evaluation highlights table on the Dashboard.
What to do next
When a model is identified as performing below the threshold that you specified, you can train a
replacement model, deploy the replacement model, and update your application endpoints to point to
the new model.