Comparison of Resampling Methods on Mobile Apps User Behavior

Mobile applications have become a vital part in modern businesses where products and services are offered in real-time. As many people have adopted to mobile apps, it is not uncommon that some of the applications are used for a few times and then abandoned. This "churning" effect on mobile apps has become a wide topic of interest among businesses to understand the factors affecting the user abandonment. This includes predicting and identifying the abandoning users before-hand to actively engage users to have more active and loyal app users. With datasets for churning, there is often a class imbalance problem where the retained user group is the minority class. We study and assess several over-sampling methods and under-sampling methods combined with several classification methods to improve the prediction ability and model performance of mobile app user retention using data available from a local mobile app developing company. The results indicate that combining under-sampling and over-sampling techniques improve overall model performance and right pick of re-sampling techniques are critical for better predictive results.