1
$\begingroup$

Let me preface the following by saying I have very little knowledge of the ins and outs of machine learning but I am familiar with statistics and programming.

I am working on a project that considers multiple variables to help predict whether an ID is "high-risk", that is, a high chance they will terminate their contract with us.

Is this something machine learning could solve? I have read and watched a few videos on Support Vector Machines and they seem to be a possible solution, but again I don't know if I'm just force-fitting SVMs into my project or if there is a more elegant solution to my problem.

Assuming ML is the correct approach, how do I go about the system recognizing a high-risk, not yet terminated ID? Do I load the data of IDs that have already terminated and once it understands the variables that contribute it can then predict for the ones that aren't yet terminated?

In my head I have this idea of the model being able to take all variables, spit out a single value indicating the likelihood of termination so that we can then target those to retain.

Examples/key words to search for/YouTube links are all appreciated. My main problem at this point is I don't know what to even search for to start gaining traction.

$\endgroup$
  • 1
    $\begingroup$ I'd be inclined to model the failure probability via logistic regression (i.e. not use it for classification as such, as an ML person might tend to, but actually predict/estimate the probability, which is after all what logistic regression is for); this would leverage your stats background better as well as more directly address what seems to be the underlying question of interest (identify which records have the highest predicted probability of failure). If you want to use fewer than all of the variables, you'd still want to be using out-of-sample model assessment for ... ctd $\endgroup$ – Glen_b Jan 14 '17 at 3:06
  • $\begingroup$ ctd.. building your model (e.g. using cross validation). If you have relatively few variables then selection may be a non-issue and you could stick with a full model, but some regularization would probably still be in order in order to get better out of sample predictive behaviour (again, methods like cross-validation can be of value there). Your mention of "high risk but not yet terminated" suggests that you're interested in risk of failure/termination over time, in which case, survival models might be relevant. Again this may better use your stats knowledge since its a kind of "regression" $\endgroup$ – Glen_b Jan 14 '17 at 3:06
  • $\begingroup$ Thanks for the reply. Logistic regression does seem to be a valid approach, I have been spending the past few days messing with the Cox Model but I don't know if it would provide me with a single value (probability). Could you apply a logistic regression to a survival model? I don't so much care about the hazard ratio that survival models provide, but I do agree that the data I have is laid out exactly like survival data, so if I could find some value to show the risk of failure given the variables over the next x months that would be ideal. $\endgroup$ – Matthew Snell Jan 14 '17 at 3:54
  • 1
    $\begingroup$ I wasn't specifically suggesting a Cox model, just survival models generally (incl. parametric survival, if it suits). If you're looking at individuals cross-sectionally, where each individual thing has a single chance of failure that depends on some set of fixed predictors that would suggest things like logistic regression. If you are following a set of individuals over time, and the longer you observe them the more will have failed (while some might never fail in the observation period), but the failure rates depend on predictors as well as the survival time, that's a survival problem. $\endgroup$ – Glen_b Jan 14 '17 at 6:01

Your Answer

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

Browse other questions tagged or ask your own question.