AI Case Study

"Paypal needed to redefine the metrics the company associated with churn, so the executive team could run KPIs and better understand the long-term health of the business. In addition, the operational teams, including marketing and product teams, needed more accurate, actionable information to help them run campaigns to retain or reactivate customers. To get the information each internal stakeholder needed quickly, Paypal’s Senior Data Scientist, Julian Bharadwaj, and his team began developing a predictive model to show when a customer would churn, or not. During this process, the team jumped straight into using random forest and GBM with H2O, running through R.

As the models were developed, Bharadwaj looked at transaction and behavioral variables as well as demographic data for customers who had churned. The first two turned out to be critical indicators of churn; the latter was not very useful, so the team dropped it. Using H2O made it fast and relatively easy to do that: the models could be modified across multiple parameters and run multiple times very quickly, so Bharadwaj could ensure the validity of the output.

Now in production, Paypal uses H2O on Hadoop to run a predictive modeling factory – large-scale, rapid modeling – that helps Paypal run more sophisticated and e ective marketing programs that reduce churn.

Immediately, Paypal began seeing great results. Initially, modeling on a laptop or a really big virtual machine took a long time when using R and ODBC on hundreds of thousands of rows of customer data. In fact, it wasn’t unusual for modeling to take around 6 hours for a subset of customers, and scoring on the entire customer base would take close to 72 hours. When Bharadwaj switched to using H2O on Hadoop, those times diminished to 10 minutes and 5 minutes respectively to train and score on Paypal’s entire customer base.

“When we started, it was a long, drawn-out process of testing,” said Bharadwaj. “We used Python, then moved over to R as a platform, and then realized that for the volume of data we had, and for the complexity of the models we were fitting, those solutions took a long time. We had to gure out ways to do it quickly, and that’s when we started exploring H2O.” Bharadwaj continued, “What took me 6-7 hours, now took me less than 30 minutes on just development hardware”."