Investigating Churn Detection in Dynamic Networks

Retaining users and customers is one of the most important challenges for the service industry from mobile communications to online gaming. As the users of these services form dynamic networks that grow in size, predicting churners becomes harder and harder. The changing behavior of users and type of services changing day by day make it di cult to monitor the mobility of customers. However, from the service providers point of view, convincing a customer to keep using their services is more e cient way than the gaining a new customer. In this thesis, I explore the use of anomaly detection for churn prediction. Due to the reason that users generate a huge amount of data during the use of services, I approach to the problem in terms of stream clustering methods. To this end, I evaluate bio-inspired and massive online data analysis techniques on public data sets, which are well known for clustering and classification tasks, as well as real world cell phone and online gaming data sets. I discuss the results of each technique from the perspective of usage of e cient features, sensitivity analysis on the parameters of the respective techniques as well as their performance.