Which App Features Are Being Used? Learning App Feature Usages from Interaction Data

In the dynamic and fast-growing app market, monitoring and understanding how past releases are actually being used is indispensable for successful app maintenance and evolution. Current app usage analytics tools either log execution events, e.g., in stack traces, or general usage information such as the app activation time, location, and device. In this paper, we focus on analyzing the usages of the single app features as described in release notes and app pages. We suggest monitoring nine app-independent, privacy-friendly interaction events for training a machine learning model to learn app feature usages. We conducted a crowdsourcing study with 55 participants who labeled 5,815 feature usages of 170 unique apps for 18 days. Our within-apps evaluation shows that we could achieve encouraging precision and recall values already with ten labeled feature usages. For certain popular features such as browse newsfeed or send an email, we achieved F1 values above 88%. Betweenapps feature learning seems feasible with F1 values of up to 86%.