Identifying features of Android apps from execution traces

Understanding a program and the features it provides is essential for a number of software engineering tasks, including refactoring, debugging, and debloating. Unfortunately, program understanding and feature identification are also extremely challenging and time consuming activities. To support developers when they perform these activities, we propose FeatureFinder, an approach that aims to identify and understand the features of a program by analyzing its executions. Specifically, we defined our approach for Android apps, given their widespread use. Given an app, FeatureFinder generates traces that capture different properties of the app executions through instrumentation. It then leverages the user events in the trace to split the trace into segments, and clusters these segments based on their characteristics, using a classifier. Each identified cluster indicates a feature exercised in the execution. Finally, FeatureFinder suitably labels each identified cluster, so as to provide a human-readable description of the corresponding feature. We performed a case study in which we used FeatureFinder to identify features in two executions of the K-9 Mail app. In the study, FeatureFinder was able to correctly identify 6 of the 11 manually identified features, which we believe is an encouraging result and motivates further research.