07-21-2015 10:05 AM
I've been tasked with creating a customer life cycle model in EM. Our customers have certain "constant" features (which country they live in, what language they speak, etc.) and then monthly data related to their use of the account.
I was wondering how will EM deal with the combination of the two data "types" (I feel like there's a better word for this but cannot think of it), that is, will having data which isn't monthly repeated monthly skew the model in any way?
I've read a few books on using EM and done some googling but I can't find an answer to this.
As a related question, I am planning to join all the data into a single table (in EG) and then bring it into EM. Would it be better to bring the constant and monthly tables in to EM and merge them there (using the merge node)?
Should I merge them at all? I know that the decision tree node can take multiple nodes as input but again I haven't been able to find any information on how it deals with having multiple nodes as input.
Thanks in advanced for your help! Let me know if there's more information I should provide or anything.
07-26-2015 03:26 PM
I think it is a standard practice to use both types of features for a predictive model. Most models use constant features like demographics or type of account with transaction summaries like the monthly in-house behavior you described.
How are you defining customer lifetime value? Expected profit in X number of months ahead?
There are many times of approaches to predict this. Most model nodes handle these two types of inputs.
How complex is your model? Is it a two-step model, for example, you predict whether a customer is going to return as a binary target, then you predict the expected value as an interval target?
Or are you predicting the lifetime value directly?
What models are you trying for each step?
07-27-2015 07:52 AM
The plan is to predict the customers monthly behaviour at a certain account age. The predicted target will either be interval or continuous I haven't decided yet.
As for the model, if I'm understanding your question correctly, I'll be using a single step model. The plan is to use a decision tree model, and then preform a logistic regression to help explain the model better.