Abstract
User intended actions are widely seen in many areas. Forecasting theseactions and taking proactive measures to optimize business outcome is a crucialstep towards sustaining the steady business growth. In this work, we focus onpre- dicting attrition, which is one of typical user intended actions.Conventional attrition predictive modeling strategies suffer a few inherentdrawbacks. To overcome these limitations, we propose a novel end-to-endlearning scheme to keep track of the evolution of attrition patterns for thepredictive modeling. It integrates user activity logs, dynamic and static userprofiles based on multi-path learning. It exploits historical user records byestablishing a decaying multi-snapshot technique. And finally it employs theprecedent user intentions via guiding them to the subsequent learningprocedure. As a result, it addresses all disadvantages of conventional methods.We evaluate our methodology on two public data repositories and one privateuser usage dataset provided by Adobe Creative Cloud. The extensive experimentsdemonstrate that it can offer the appealing performance in comparison withseveral existing approaches as rated by different popular metrics. Furthermore,we introduce an advanced interpretation and visualization strategy toeffectively characterize the periodicity of user activity logs. It can help topinpoint important factors that are critical to user attrition and retentionand thus suggests actionable improvement targets for business practice. Ourwork will provide useful insights into the prediction and elucidation of otheruser intended actions as well.