RESUMO
In recent years, the use of mobile applications for digital service
is being widely deployed in a varied range of contexts. With this,
predicting the possibility of churn is vital for selecting users that
could be targeted with user-retention campaigns. This technique
is commonly referred as Churn Prediction Problem (CPP). Most
studies in the literature use traditional machine learning techniques
to predict churn, and neglect the users’ privacy. In this work, we
propose a privacy-preserving solution that uses neural network to
predict churn of mobile services. Our solution, called AutoRGNN,
requires only the installation and uninstallation sequences of mobile
apps, and integrates Recurrent and Graph Neural Networks. In
comparison with a traditional baseline approach in a large-scale
and real scenario, AutoRGNN was capable to increase the recall
and precision up to 19% and 7%, respectively.