Nia McCash
Mega Sage
Mega Sage

Since the New York release, ServiceNow’s Predictive Intelligence product provides Clustering capabilities. But what do you do with it? 

In this article, I will share a couple of example use cases for leveraging the clustering framework. 

Dan Grady writes: “Clustering is used in more exploratory fashion that other frameworks… using the output of the training to identify opportunities - whether it be additional automation through [Integration hub] or maybe the opportunity to role out a handful of Virtual Agent conversations aligned to the top clusters of tickets coming in.” 

Mark Roethof also writes: “With Clustering for example, you could - within minutes of work - analyze incidents, etc. This could give you insight in great subjects to cover with Virtual Agent.” 

Both experts have provided use cases for making use of the Clustering framework which I break down here...

Identifying Top topics to be addressed by Virtual Agent (VA)

Often, organizations want their VA to help deflect some of the most frequent issues or questions. But identifying the most frequent issues or questions is not always straightforward. If you are unsure of what topics you should address with VA, or perhaps you have some idea but want data to backup your hypothesis, the clustering framework can help.

As an aside, you can also use Clustering to help identify high-impact KB articles to write, but ServiceNow provides a FULL solution chain for that called Knowledge Demand Insights. More on that from my article Understanding Knowledge Demand Insights.

 

As for clustering specifically, the docs site provide instructions on how to create and train a clustering solution.

Depending on your data and/or your clustering configuration, your initial solution might look a little like the image from the docs site.

Rectangle divided into smaller labeled rectangle regions of different sizes and different shades of teal

The lighter colours indicate cluster qualities of about 50-75. The higher the quality, the more alike the records are in the cluster.

For example, if you get a cluster for “password reset” that is of low quality, you may find that not all of the tickets identified in that cluster are really about password resets. 

Tuning Your Clustering Solution

To get better clustering results, you need to experiment with the solution definition. Refining the parameters for your clustering solution may take a bit of trial and error, and that’s ok! 

I found a few tweaks helped in my situation: 

  1. Adding my own list of stop words - to filter out words like please, thank you, hello, and, in my case, location information like room numbers or building names that were not significant for my use case.

    Some of these stop words become apparent after you train your solution the first time and explore the results

    A list input field labeled 'Stopwords' with the values 'Default English Stopwords' and 'Common Words to filter' in the input field

  2. Increasing the Minimum number of records per cluster.
     
  3. Adding Purity Fields to further determine the quality of your clusters. More on Purity Fields below.

Note that as you tweak your solution definition, the coverage might decrease.  For example, if your filter identified 10,000 records to be clustered, only 3,000 of those may end up in a cluster. 

This is ok. Remember the goal is to identify to the top issues or questions to address, not to find a solution that is applicable for all records.  

Tuning the above parameters, I was able to find higher quality clusters. Note the Purity Fields:

A pop up which shows Purity baesd on subcategory, category, and assignment group - all at 99%

This brings me to the 2nd use case for the Clustering Framework as mentioned by the experts quoted above...

Exploration and analysis to identify opportunities

Here is just one example…

One of the things I noticed when going through my clusters is that some of the Purity Fields had relatively low percentages like this one: 

A pop up showing purity fields of subcategory and category at 43% while assignment group is at 99%

The above tells me that the category and subcategories vary for records in this cluster. If all the records are about the same thing or very similar things, it’s possible that: 

  1. The categorization structure is poorly designed (eg. does a ticket about someone having problems sending email via Microsoft Outlook belong in the Email category or the Software category), or…
     
  2. The fulfillers are not categorizing tickets correctly. There could be an opportunity here for further training and improvement. 

 

Last but not least, I should mention there is a ServiceNow share app for Clustering Recommendations Utility:

This Utility simplifies the Clustering analysis based on Predictive Intelligence. You can add recommendations and capabilities to the clusters in order to identify automation use cases.

This Utility enables you to calculate the potential benefits based on Volume Deflection, Manual Effort Reduction and Productivity Improvement for Incidents, Requests, Catalog Tasks, Security Incidents, Cases, HR Cases, Calls, and Interactions.

 

Hope this article has been helpful in providing ideas and insight on using Predictive Intelligence’s Clustering framework. If you found it helpful, please mark helpful below and/or leave a comment.

Version history
Last update:
‎03-09-2021 05:30 AM
Updated by: