How to Do Prescriptive Analytics in Dakaiku

Solved!
sdfungayi
Level 3
How to Do Prescriptive Analytics in Dakaiku

How do I implement prescriptive analytics in DSS?
My churn prediction model tells me which employees are likely to leave. I now want DSS to suggest to me the best course of action to stop them from leaving, such as which variables to adjust and by how much.
How can I do this in DSS?

0 Kudos
2 Solutions
tgb417

@sdfungayi 

I've been thinking about the same type of questions.  Once I have a model how do I operationalize the model?  It sounds to me like you are looking for insights at the employee level about which features are most likely to lead to churn. 

I've been considering using a feature of the Scoring Visual recipe that produces an explanation of which variables were most important to the specific prediction for that specific record.

Compute Individual Explenations.Compute Individual Explenations.

 

This will produce an Explanation column in your output.  With JSON listing the top N explanations features.

Details about Scoring.jpg

 

I'm thinking that this would make a great diagnostic around why a particular employee would churn based on the validity of the model you built.  

You might then have to build a further optimization model to look at what treatments you have, and the implication of the likely reduction in Churn, and then optimize for what you value in these treatments.  (For example is cost more important, than retention.  Do you care for how long you have employee retention...)

This may be a start for your thinking about this and some other features of DSS.

However, that said, because you are talking about a potentially important analysis related to employee churn. Where there are likely legal/ethical consequences of using a model that is put into production.  I invite you to find folks with more experience on this particular type of model than I do. To advise you on the best way to produce such a model.  There are lots of things that could get you down a bad track like biased data, problematic feature inclusion, detection of protected characteristics like age, gender, ethnic and cultural background, without having a column with such data.  Then there is the question about what you are optimizing for, and I'm sure that the list goes on, and on, and on...

--Tom

View solution in original post

sdfungayi
Level 3
Author

Sounds like a good place to start. Thanks @tgb417.

View solution in original post

0 Kudos
3 Replies
Liev
Dataiker Alumni

Hi @sdfungayi 

DSS allows you to structure your projects in flexible ways, and I suggest that you look at your internal set up in order to create a meaningful architecture that fits into this setup. 

In the scenario you describe, you may have a project that is in charge of predicting churn. How those results are consumed or acted upon needs alignment within the organisation. For example, you could expose your churn model as part of a real-time API, or perhaps an email to a department with the results. 

I recommend you start from an ideal scenario of how such information would be consumed or acted upon and build backwards.

Good luck!

0 Kudos
tgb417

@sdfungayi 

I've been thinking about the same type of questions.  Once I have a model how do I operationalize the model?  It sounds to me like you are looking for insights at the employee level about which features are most likely to lead to churn. 

I've been considering using a feature of the Scoring Visual recipe that produces an explanation of which variables were most important to the specific prediction for that specific record.

Compute Individual Explenations.Compute Individual Explenations.

 

This will produce an Explanation column in your output.  With JSON listing the top N explanations features.

Details about Scoring.jpg

 

I'm thinking that this would make a great diagnostic around why a particular employee would churn based on the validity of the model you built.  

You might then have to build a further optimization model to look at what treatments you have, and the implication of the likely reduction in Churn, and then optimize for what you value in these treatments.  (For example is cost more important, than retention.  Do you care for how long you have employee retention...)

This may be a start for your thinking about this and some other features of DSS.

However, that said, because you are talking about a potentially important analysis related to employee churn. Where there are likely legal/ethical consequences of using a model that is put into production.  I invite you to find folks with more experience on this particular type of model than I do. To advise you on the best way to produce such a model.  There are lots of things that could get you down a bad track like biased data, problematic feature inclusion, detection of protected characteristics like age, gender, ethnic and cultural background, without having a column with such data.  Then there is the question about what you are optimizing for, and I'm sure that the list goes on, and on, and on...

--Tom
sdfungayi
Level 3
Author

Sounds like a good place to start. Thanks @tgb417.

0 Kudos