What can AI teach us about behaviours leading to churn?
Readers First Initiative Blog | 05 July 2020
Leading publishers invest in data science to predict who’s going to churn so they can prevent it. Advanced analytics can be used to inform tactics as well as the engagement strategy of news media.
Propensity modelling was the focus of the June 30 class by Michal Ciesielczyk, head of AI engineering at Deep.BI, an analytics platform provider, at the recent INMA Master Class on Digital Subscriber Retention (you can still catch up with the class on-demand).
In an interview with INMA, Ciesielczyk presented the average importance of features indicated by two models trained on data of mid-sized national news publishers in Europe and in Asia. His team used an ensemble of machine learning algorithms.
- The most predictive was the time since the first visit ever by a reader, or the tenure, measured by days and then translated into a loyalty score.
- Then came temporal variables, such as long-term change in engagement measured by the RFV score (measures recency, frequency of visits and volume of articles read), visit frequency, time since the last visit, short-term change in RFV, and the RFV score itself.
- Other features found predictive were total attention time spent interacting with the site, whether the reader subscribed to a newsletter or shared articles on social media, and the number of articles read or images clicked.
The models of Deep.BI were trained using mostly Web site behavioural data. Based on INMA interviews, other vendors and publishers use also features such as payment data, information about content, marketing channels, and pricing.
- For example, Dagens Nyheter in Sweden identified the subscriber’s tenure, acquisition channel, and payment method strongly predictive.
- Apart of temporal variables, The New York Times found visits to the home page mattered as well as the type of content consumed.
Advanced analytics can be used also to help answer other questions across the customer journey, such as who’s going to register, what content or features best to recommend next, and what offer best to show, when and using what channels.
Interested in an introduction to the Ph.D.-level analytics? Michal Ciesielczyk of Deep.BI delivered a primer on propensity modelling last week at the INMA Master Class on Digital Subscriber Retention. Buy a ticket to catch up on-demand.