I would like to use users' consecutive inactivity time in order to predict probability for not returning.

For example, I would like to be able to say that if a user was inactive for a month, the chance he returns (likelihood would be correct here?) is 40%. A user that was inactive for 2 months, the chance he returns is 5%. OR the opposite where according to the inactivity, the chance they die is xx%.

The data that is available to me is obviously all the uses, and from that I could extract all existing inactivity intervals.

If I understand it correctly it is the opposite of survival analysis where I use time a live as predictor of death probability. But perhaps Survival analysis could be used in some different form?

If in your answer you could add the way to use your suggested model in R it would be great!

  • $\begingroup$ I guess it might be easier to obtain data if you timebox the returning probqbility? like eg '40% chance of returning in next 3 months'? $\endgroup$ – Hugh Perkins Oct 9 '17 at 12:48
  • $\begingroup$ given that he was inactive x number of days? I'm not sure it is simpler as I don't want to predict within which timeframe he will return - and subsiquantely probability of return in many different timeframe: 20% in the next month, 30% in the next 2 months etc' - but the probability of return (or death) period. $\endgroup$ – Kantushov Herman Oct 9 '17 at 14:13
  • $\begingroup$ well, it will be easier to get training data on. I'm not sure how you can get training data for someone never returning, since you'd need to wait until their death to know for sure? $\endgroup$ – Hugh Perkins Oct 9 '17 at 14:19
  • $\begingroup$ you're right. So let's assume that if a user haven't return for 45 days he is considered dead. $\endgroup$ – Kantushov Herman Oct 9 '17 at 14:23
  • 1
    $\begingroup$ Label is: returned/didnt return, in the 45 day window. Its a binary classification problem. Input data is all the features you have. Then, you can use any classification model of your choice, eg logistic regression, neural nets, etc. Note that for any particular example/person, you can create a ton of different examples, by sliding the 'prediction date' backwards and forwards. $\endgroup$ – Hugh Perkins Oct 9 '17 at 14:58

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.