Machine Learning for Incident Classification? Where do we start?

[barryrailton]

Hi guys, we are interested in developing a machine learning algorithm that can auto assign the incidents to the correct groups/teams. I see this: https://hi.service-now.com/kb_view.do?sysparm_article=KB0597625 Where can we start?

[dawidkubiak]

Hi Barry,

I will write what I think can help you:

1) You need table to contain keywords and group/teams. This table will be populated for example when incident is closed, with keywords and correct assignment group. (so incorrectly assigned wont be a problem)

2) Logic for keywords can be simply as take for example whole description and split it by "space". Then for each row in explained above table you can search each word from description. So you will have number of matches for each row which will decide to which group it can be assigned.

3) Each word not matched can be added to table from 1) as temporary and when temporary word hits certain number of entries it can be moved to field where words are matched.

Keywords                                                                     | Assigned Group     | Temporary keywords

keyword1, keyword2, keyword3 | Group 1                             | tempkeyword1 - 2, tempkeyword2 - 1

You can use blacklist of words which you dont want to use and any other solution. Fields in table can be as strings or references to other tables.

For filling table with historical incidents data, you can get all description words and enter them into table with proper group in one time run scheduled script job.

There can be more specific logic applied, my solutions is simply but one thing to remember - remove duplicates before comparing

Happy to hear opinions and proposals from others

Dawid

[barryrailton]

Hi, thanks for the reply, Dawid. I was thinking more of a machine learning solution using the whole ticket, all fields (aka "features"). This is similar to credit card fraud or spam filtering using Bayes Theorem where you use every piece of info available and then apply linear regression. But beyond that can the algorithm run on the same cluster as the DB and will we need Hadoop or something similar?

[dawidkubiak]

You have great idea for fully use machine learning but I wonder if ServiceNow is in that stage yet. There is almost no interaction where scripts run on platform, there are two nodes for instance which picks actions and executes them. If any script takes long time to execute (really long timerun with many queries) it triggers incident on ServiceNow side and they will contact you with question if they should terminate this transaction. Also such transaction will block other which are on same node queue because many things run behind on async mode. I talked with people which were interested in BigData and they are using Azure platform for BigData with integration to ServiceNow because performance issues were occurring even without such analysis.  

I am not expert on this topic but from my experience such analysis depends on quantity of data which will be used. I would suggest that in current state it could support simple algorithms but its not ready now for complex solutions which requires modeling of data with many iterations.

ServiceNow recently acquired company which specializes in intelligent automation so in future we could be in completely different position.

ServiceNow Acquires DxContinuum

[sergeyfomin]

Hi Barry,

You can try our solution for auto assign the incidents to the correct groups/teams. See site 1mind.ai . We made this solution for internal using, but it seems to me that the solution benefit many companies. Our customer service team using that solution for auto assign fields, like urgency, category, assignment_group and so on.   You can use whole ticket (any fields) for auto assign correct group.

BR,

Sergey.