Reflection

Abstract

By knowing which upcoming action a user might perform, a mobile application can optimize its user interface for accomplishing the task. However, it is technically challenging for developers to implement event prediction in their own application. We created Reflection, an on-device service that answers queries from a mobile application regarding which actions the user is likely to perform at a given time. Any application can register itself and communicate with Reflection via a simple API. Reflection continuously learns a prediction model for each application based on its evolving event history. It employs a novel method for prediction by 1) combining multiple well-designed predictors with an online learning method, and 2) capturing event patterns not only within but also across registered applications--only possible as an infrastructure solution. We evaluated Reflection with two sets of large-scale, in situ mobile event logs, which showed our infrastructure approach is feasible.