Schibsted’s subscription model predicts purchases, drives conversions with targeting

By Scott M. Scher, Columbia University

By using machine learning and AI Schibsted is able to predict likely buyers. This project is based on three key ideas. One, business impact. Two, data science leadership. Three, multi disciplinary team driving results. They used this project to see how user behavior on the site relates to purchasing and to see if they could use that to predict potential buyers.  In 2018 they had 175,000 users and used real-time data tools to evaluate user behavior related to buying.

The project had two phases, proof of concept phase where they focused on proving the solution was useful to business, that business was willing to use the solution, and collaborating with the commercial teams. They had regular meetings together to explain what they were doing, the kind of signals they were seeing in the data, and how they were approaching the problem. This created a lot of understanding, buy-in, and trust between the teams.

As a result the business was committed to using the solution once it was made. The basic problem they were addressing was figuring out from a particular period of observed user behavior on the website how likely are they to purchase based on that behavior. The observed behavior was how many pages a user views and when they viewed and the target period was which of those users did and did not purchase subscriptions.

In the second phase, targeting. They use a random Forrest algorithm to make predictions of behavior-based on past data. With machine learning algorithms identify who will purchase can be automatically determined from behavior data. They were able to determine that users who were most likely to by were those that visited the site most recently, were weekend visitors, who  visited the site via multiple sources, that accessed the site from multiple devices, that spent more days coming to the site, and that read more articles once on the site.

The way this is used is through a scoring matrix. You take all the users and give them a score on how likely they are to purchase and than you simple order them from most likely to least likely to purchase and than from the historical information in the data set you look for how many are in the top ten percent of most likely to purchase. This means that looking at a random sample of ten percent of users you should get ten percent of conversion. The actually results of the model is upwards of 50 percent, proving that the model is good at determining purchasing behavior.