Creating a Prediction Classifier

Classifiers are Natural Language Classifiers that are used to predict the value of fields based on the input of a string field. They are also used to define tables used by Sofi Intelligent Search.

A Prediction Classifier is trained using historical records from a ServiceNow Table and then returns predictions (Classification fields) for input text (Search fields). A Classifier will have one of more Classifier Contexts which control the behaviour of the classifier in different contexts (eg via UI, email or Virtual Agent). Each Classifier is triggered by a Client Script.

Defining a Classifier

Perform the following steps to create or modify a prediction classifier.

  1. Navigate to Sofi > Classifiers and Select New
  2. Configure the following fields:

πŸ‘

The mandatory fields that must be configured are:

Name: Unique name of Classifier
Table: Which table will the Classifier be applied
Type: For prediction Classifiers the Type will be "Classification"
Search Fields: The text field used to trigger classifier
Classification fields: The fields on the form / table that will be predicted by the classifier

The table below provides a list of all of the parameters on a Classifier Record.

Field

Description

Notes

Name

Unique name of the Classifier

eg Incident

Table

Table the classifier reads data from

eg incident

Status

Status values of the classifier:

Read Only field. Primarily to indicate the status of training

Type

Classifiers are used for Intelligent Prediction, Search and Response.

For Prediction Classifiers, set Type - Classification

Link Knowledge

Enables the Knowledge Attach Button on the Knowledge Search Classifier.

Use HTML: Copies the Knowledge Article as HTML into the Knowledge Text field
Use Reference: Copies the Knowledge Article reference number into the Knowledge Reference field

Knowledge Text

Defines the field on the target Table form (eg Incident form) that the Knowledge Article's text field will be inserted if Link Knowledge is active and the Attach button is selected on the Knowledge Search Classifier result.

Eg Work Notes

Knowledge Reference

Sets the Knowledge record in a field on the form. Field must be a Reference field with target table Knowledge.

Description

Description of the classifier

Restriction

Use this condition builder to restriction / define the source records that will be used to training the Classifier. For example, you may only want to use Records that were created in the last 12 months.

Typically there is no need to restrict the training data.

Search fields

The text field used to trigger classifier

Note: Normally on single fields selected

Classification fields

The fields on the form / table that will be predicted by the classifier

eg Category, Subcategory and Assignment Group

Training Start

The date / time that the Classifier most recent training started.

Training End

The date / time that the Classifier most recent training started.

Advanced Configuration

The following section details some of the advanced configuration options for a Classifier.

Sofi Classier Advanced Configuration

Classifier Filters

Classifier Filters allow you to define criteria for improving the relevancy of the classifier results. A classifier filter uses information on the current record, such as Location, to restrict the training data taken into account by the classifier.

Field

Description

Notes

Filter field

Classifier uses these fields on the record to filter the results

For example, Assignment groups for each state (e.g. QLD OPERATIONS)
The Filter Field can be set to 'assignment_group.location.state'. The Classifier returns then these records and can filter out Assignment groups from other states.

Filter source fields

Classifier uses the Filter source fields to identify the data source used by the Filter field.

For example:
Classifier depends on the state (location) of the caller. The Filter source field identifies where the state field is defined. In this case the state field is located on the User's profile ==> location ==> state (caller.location.state).

Modifying Classifier Results Layout

Field

Description

Notes

Layout

Provide the option to override the default layout of our Classifier Prediction results are presented in the Sofi Service Agent Assistant -
Intelligent Prediction UI sidebar

By default this field is blank and Sofi uses a default UI layout for Prediction Classifier results.

Generate Default Template

A UI Action at the bottom of the Classifier form that generates a Layout for the configured Classification fields

You can use this field to modify the layout for the Sofi Service Agent Assistant -
Intelligent Prediction UI sidebar

An example of the Layout syntax is shown below:

<LABEL>Classification:</LABEL>
<SECTION>
<FIELD-L2>{{ fields.category_display }}</FIELD-L2>
<FIELD-L2>{{ fields.subcategory_display }}</FIELD-L2>
</SECTION>
<LABEL>Assignment:</LABEL>
<SECTION>
<FIELD>{{ fields.assignment_group_display }}</FIELD>
</SECTION>

Prediction Classifier Layout on SAA User Interface

Other fields

Field

Description

Notes

Data analytic fields

Used to collect additional information for data analysis.

Internal use only, please leave blank.

Request parameters

Internal use only, please leave blank.

Result processor

Internal use only, please leave blank.

Training a Prediction Classifier

Before Sofi can be used we need train the models, this is done by navigating to the ServiceNow menu Sofi > Configuration > Classifiers. Select the Classifier and then click on the Train Classifier link at the bottom of the form.

ServiceNow will request confirmation to start the training process, select Ok to start. The duration depends on a number of factors and can take between 20 minutes and a few hours. The Classifier provides a Status field indicating the current state allowing you to confirm when training is complete. Once completed, the Training End field will be populated. Note that during Training, the classifier will still make predictions however the results may be inconsistent.

πŸ“˜

Estimated Training Duration

As a guide, 100,000 may take 20 mins. 500,000 may take 2 hours. 1,000,000 records may take 4-5 hours.