1
$\begingroup$

Could you explain the need of temporal abstraction in churn analysis intuitively with a simple example? I tried Google but there are not any clear answers , especially for churn analysis.

$\endgroup$
2
$\begingroup$

AFAIK, churn analysis is the domain specific lingo of survival analysis applied to customer relationship management. The literature and google-ability of "survival analysis" is far superior to "churn analysis."

A popular way of modelling survival is the Cox proportional hazards model. This model assumes that the influence of an event on another event is independent in time.

For instance, suppose you hypothesize that sending an thank-you email hours before a customer is offered to renew his/her subscription has a greater effect than days before. This violates the proportional hazard assumption and you need to extend your model with time varying covariates.

This modified model is termed the Extended Cox Model. You can perform a Wald test to see if this additional explanatory variable better describes the data.

$\endgroup$
4
$\begingroup$

Temporal abstraction is fancy pants jargon for changing the reference point of a time series variable. It is a type of variable transformation.

A common example is that you have behavioral data that is indexed by calendar time, like number of minutes used in a given month on a cellular plan. This is typically how this sort of data is stored in a database. You can change the reference point to be time before churning takes place, or perhaps the current month if the customer has not churned. This often allows patterns to emerge that predict churn. For instance, a customer might slow down how quickly she ships back Netflix DVDs in the three months before terminating the subscription service completely. You can use that information to intervene somehow.

Here's graphical example, where the colors correspond to behavior and T-1 is the month before churning:

enter image description here

Source: Handbook of Statistical Analysis and Data Mining Applications, p.341.

$\endgroup$
  • $\begingroup$ Thanks!So, why can't the original variables be used to predict churn? Where does the difficulty comes in to model such data? $\endgroup$ – leviathan Aug 7 '14 at 23:15
  • $\begingroup$ One reason is that you have more data to work with, so the pattern is amplified. Suppose I use the July data to predict churn in August. I have 4 data points to work with (rows 1,2,4 & 5). Yellow leads to churn with certainty. Blue and green seem equivalent (no churn for either one), but it probably makes sense to give the coupon to the green. That is much more apparent on the right to my eye. $\endgroup$ – Dimitriy V. Masterov Aug 8 '14 at 0:10
  • $\begingroup$ Another example might be seasonality. Suppose people take a long time to return the DVDs for 2 reasons. In the summer, they spend more time outside, so it takes them longer to watch the films. In the rest of the year, taking longer means you loose interest and will churn. The behavior lacks signal in the summer, but signals something elsewhere. The conclusion will depend on when you do the analysis if you use the method on the left. $\endgroup$ – Dimitriy V. Masterov Aug 8 '14 at 0:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.