I have built a customer engagement model using logistic regression, The approach was to take a group of customers we consider engaged, then build model to find lookalike in the customer base.
The most important variables were as you will expect, number of visits in the last 12 weeks, number of distinct product they use in the last 12 weeks etc...I have also entered variables, like visits in the last 1 week, last 4 weeks etc but the most important were only 12 weeks var.
My question is if there is any change , like price increase of products or any other events that will impact the purchase. How do I make sure in the model that it won't take 12 weeks to see how customers are getting disengaged etc..
It sounds like you are doing some type of matched-pairs design which is not typical for a data mining problem which routinely analyze a huge number of observations. The Memory Based Reasoning node (MBR) identifies neighbors about each observation that might be useful in what you are describing.
Regarding the 12 week vars, if you are saying there is no information in the shorter term variables, it is possible you won't see the change except in the 12 week vars, but you can always update the 12 week vars each week (or every 4 weeks if you have monthly data). Using rolling windows to refresh your variables (not done automatically by SAS Enterprise Miner) allows you to update your data more regularly so that you don't have to wait to see the change. In general, check the performance of the scored data against the training data and refit when necessary.
Hope this helps!
Doug
