Predictive Intelligence - Regression framework issue with duration type output

[Istv_n B_]

Hello,

I am currently trying to setup a solution with the regression framework, which would predict the business resolve time field on the incident table based on the inputs. I have tried tried with varying input fields and different filters but it always runs into the same error, which is: 

SE0035:Training terminated due to Exception. SE0075:Solution training failed as either the data used is not sufficient or the input field(s) is not predictive of the output field.

The input fields I have tried with: Assignment Group, Business Application, Short Description, Description, and combination of these.

What I have noticed is that it only happens, when I am trying to work with a duration type of outputs. For any other numeric fields, it works perfectly. Based on the documentation, it should work for duration types as well. 

Anyone experiencing the same?

Any workarounds / ideas are welcome. 🙂 

Thanks!

[Brian Bakker · Accepted solution]

Hello István,

Indeed, the error indicates that the data in the Input fields Assignment Group, Business Application, Short Description, Description that you have provided in your Regression solution to determine the Business Resolve Time is not sufficient or predictive of the Business Resolve Time and so it can't generate a Regression solution with the data that has been provided.

First, I suggest trying only with Short Description and Assignment Group as Input fields. Ensure that the Input/Output fields have no empty values in the training dataset by adding this condition for these fields to the filter on the Regression solution definition. You should avoid providing empty values in the training data for machine learning, although the ML Engine can handle some empty values.

Generally, when trying to predict a numerical value with a Regression solution, you need to look at the range of values for the Output field and exclude the extreme values from the training dataset that may require a different Regression solution. For example, if you have incidents that take less than 1 minute to resolve and other incidents that take months to resolve, then the range is too extreme to make a prediction. Remove the extreme samples in your training dataset, before you submit it for training.

If successful with the Assignment Group and Short Description fields with the extreme samples removed, you can build on the solution and try adding further Input fields. However, you would need to provide more examples and we support up to 300k records for the training dataset to ensure it can generate a Regression solution.

Best regards,

Brian

[Istv_n B_]

Hi Brian!

Thanks a lot for the tips! Filtering out the extreme values worked.

[jeevan2]

Hi @Brian Bakker ,

A quick question, I was trying to predict actual resolution time in an HR Case but the value gets stored in ETTR in days. Is that an expected behaviour?

Can you please share your views.

Thanks,

Jeevan

[Brian Bakker]

Hello Jeevan,

Indeed, this is expected behaviour. We have enhanced the HR PIWB (Predictive Intelligence Workbench) in Tokyo and added the Regression solution template "Recommend estimated time to resolve". The ETTR model has the following output -

• Max ETTR: Indicates the maximum estimated resolution time in days for case completion. Estimate is based on the time taken for resolving similar HR cases.
• Min ETTR: Indicates the minimum estimated resolution time in days for case completion. Estimate is based on the time taken for resolving similar HR cases.
• ETTR in days: Stores the point estimated resolution time in days for case completion.

Out-of-box, you will be able to see the Estimated time to resolve value from:

• Employee portal view and Now Mobile view
• Agent workspace
• Platform view with a list of cases
• Case configuration view for admins

I hope this helps.

Regards,

Brian