Faculty 1: Award for the best bachelor thesis 2021

Max Bergmann, B.Sc. receives the award for his work "A Machine Learning Approach For Churn Prediction Within The Social Network Yodel". Supervisors were: Prof. Dr. rer. nat. Oliver Hohlfeld and M.Sc. Jens Helge Reelfs, Department of Computer Networks and Communication Systems.

German title of the paper written in English "A machine learning approach to predict churn within the social network Yodel".

In recent decades, the number of social networks has grown rapidly. The competition to retain customers in order to grow the platform and increase profitability is increasing. For this reason, companies need to identify potential churn in order to retain them. The problem of predicting user lifetime, churning users, and reasons for churn can be addressed using machine learning. The goal of this bachelor thesis is to build machine learning models for predicting user churn and user lifetime within the social network Yodel, a location-based anonymous messaging application for Android and iOS.

To obtain the best possible prediction results, we started with an extensive literature review, whose approaches we tested and added to a machine learning pipeline to build prediction models. Using these models, we examined performance after different observation time windows and finally compared the strongest models to identify similarities and understand the insights into learning their behavior.

The results of this work are machine learning models for a selected representative group of communities of different sizes in the Kingdom of Saudi Arabia and a country model using all data. These models are used for a regression task that predicts the lifetime of a user, as well as for a multi-label classification of a user into six different churn classes. In addition, we have also built models for a binary classification where the model predicts whether or not the user will churn within a given time. These models have shown generally strong predictive power, which decreases as the observation time window is narrowed. In particular, the binary classification showed a high accuracy of over 99%.

The best models were used to predict user churn in other communities to identify communities with potentially similar behavior. These similarities were then determined based on feature importance, with the most important features regressed on empiricism. This has shown statistically significant differences between user groups with different activity time, but to date, no clear trends have been identified that would lead us to define community behaviors.

As competition among social networks continues to grow, user engagement will remain a key marketing strategy that must be addressed through machine learning and artificial intelligence. The models created could be useful for predicting churning users within the Yodel platform to identify customers who will churn within a given time.

Anonymous and location-based messaging have received little attention in research. Therefore, the results of this work on the anonymous messaging application Yodel open up a variety of possible future tasks in this context.


Susett Tanneberger
Stabsstelle Kommunikation und Marketing
T +49 (0) 355 69-3126
Max Bergmann (Photo: privat)