Discovery of Real World Context Event Patterns for Smartphone Devices Using Conditional Random Fields

Mobile applications are Event Driven Systems that react to user events and context events (e.g. changes in network connectivity, battery level, etc.) The large number of context events complicate the testing process. Context events may modify several context variables (e.g. screen orientation, connectivity status, etc.) that affect the behavior of an application. This work examines a data set of real-world context changes on Android phones. We collect every context event that occurs on the mobile devices of 58 Android users over 30 days to identify complex relationships and patterns. This work uses Machine Learning (ML) techniques including Conditional Random Fields (CRFs) and Deep Neural Networks (DNNs) to predict sequence labels for context events. These techniques are compared to Majority Baseline (MB). The trade-offs among these methods reveal that CRF is the most effective technique for sequence prediction/labeling of the data-set. Future work may apply the data collection strategy and ML techniques to domains for emerging technologies in areas such as Internet of Thing, smartwatches, and autonomous vehicles.