Working with Predictive Scoring
Analytics Manager can deliver predictive data analytics for analysis models. It scores data based on a pre-trained data model and a dataset from which you want to predict an outcome. It attempts to make predictions about a given outcome (goal variable) in the data being processed. It can return scoring values in both human-readable and normalized form. For more information about predictive scoring, see Predictive Scoring.
You can perform predictive scoring in Analytics Manager by using one of the following methods:
• Analytics Server Connector and the prediction service in the Analytics Server
For more information, see How Predictive Scoring Using the Analytics Server Connector Works
• ThingPredictor agent
For more information, see How Predictive Scoring Using ThingPredictor Works
 |
ThingPredictor software media is no longer available for download as of 8.3.
Any new model that you publish from Analytics Builder is uploaded to the Analytics Server instead of ThingPredictor.
|
While working with predictive services, you must provide the following information that helps in scoring:
• Goal Name
• Causal Technique
• Important Field Count (only for non-categorical goals)
Goal Name
You must specify the goal for which Analytics Manager performs predictive scoring. In a predictive model .xml file, the MiningSchema tag lists all the fields in the model. The field that is specified with usageType="predicted" is the goal name that should be used when requesting a prediction.
|
|
Note the following points:
• For predictive models that were developed using PMML version 4.2 and earlier, the field that is specified with usageType="predicted" is the goal name that should be used when requesting a prediction.
• For predictive models that were developed using PMML version 4.3, the field that is specified with usageType="target" is the goal name that should be used when requesting a prediction.
|
Causal Techniques
Predictive services use the following techniques to perform predictions about a particular goal:
• Full Range — Measures the distance between the minimum and maximum score values produced by each field. This is the default technique.
• Distance from Minimum — Measures the distance between the original and the minimum score values returned by each field.
• Distance from Maximum — Measures the distance between the original and the maximum score values produced by each field.
For more information about causal techniques, see the PredictionMicroserver services at PredictionMicroserver.
Important Field Count
When you submit a predictive scoring request, you can specify the number of the important fields that influence the goal in the results. Predictive services calculate which fields have the most influence on the predictive data scores. The results include a list of fields and their associated weights for each field. These weighted fields indicate how influential each field attribute is in determining the predictive score for a specified goal.
For example, if you specify the important field count as three, the causal scoring output includes the three most influential fields and their percentage weights influence on the score.
|
|
You need to specify the important field count only for non-categorical models.
|