I have one business problem in hand which is to predict if a user will revisit the website or not within 6 months. I need to majorly understand what are the factors which make the user return and also need to give business recommendations on what can be done to make a new user return to the website. My initial idea was to do logistic regression. Lately, I read about survival analysis. I want to know if I can use survival analysis for this problem. Also, my dataset has 20k users; each user having multiple transactions; the target variable was not given I aggregated the dataset to one record per user and did some feature engineering to come up with a target variable. If I want to use survival analysis in this problem, shall I consider only the last transaction of each user or shall I use the aggregated dataset?
-
$\begingroup$ I'm not expert in survival analysis but I thought it was useful only for studying a population over time, so if I'm right that doesn't seem a good fit for predicting whether a particular user will revisit. I might be wrong though. $\endgroup$– ErwanNov 3, 2019 at 23:44
1 Answer
If you want to use survival analysis (which can be more flexible and insightful), I'd recommend this package and this great tutorial. Speaking shortly, as a result, you'll get "probability of being alive" for each customer.
If you want to use logistic regression I think it's trickier. Why I think so - Like any other churn problem, it's hard to define it properly. The definition depends on your task and where the model outcome will be used. Let's say churn is a particular amount of inactivity, e.g. 30 days. You can do an initial analysis of how to find this number. Just pick a particular date (you can do it multiple times) and check % of people who made next transaction. Important thing is - your time period from both sides should be same for all users:
- if it's a new users he's not able to be inactive for a long period, right?
- if it's a last date in the dataset (e.g. yesterday) - not all users are not able to perform transaction within 1 day. And you'll get high churn rates. So be attentive to dates.
So you need to understand from your data - which inactivity is "normal" for the average user and define it as N. After that you can label users binary like "if inactivity > N then "churned (1)" else "not churned (0)"". And you can use this label with any classification model.
