This is code for my 2 week capstone project at Zipfian Academy.
I analysed customer churn in a mobile news app.
See presentation:
I found that the company's user engagement metric didn't correspond to retention. I suggested a new metric and implemented it with the developers. See code/model/events_that_retain and code/model/eda_km_curves
I also found which news stories make first time customers come back. Moving from first to second visit was the biggest leak in the funnel at 50%. See code/model/stories_that_retain
For exploratory data analysis I drew lots of Kaplan-Meier survival curves. To find events that retain I used random forest's feature importance. It gave the insight about problems with the old user engagement metric.
I used Bayesian hypothesis testing to find news stories with highest retention.
The dataset was 8m events from Mixpanel json API imported into a Redshift SQL server. See code/mixpanel_import
I wrote unittests for the SQL feature engineering as I was writing queries following TDD. See tests/
Data analysis using python sklearn and lifelines package.
Some user sessions were broken into two or three pieces because of Mixpanel setup. See my blog post on how to avoid it and get consistent user tracking. (Most interesting part is that there is no way to export aliasing events out of MP.)