A Pervasive framework for identifying activity patterns of mobile users and predicting activities

Abstract

Smartphones has become one of the most used devices in day to day life. Even though they already have so many features, they still lack the ability to identify user’s context and the intentions. This is important for improving user experience and make existing mobile application more user friendly. The issue is that there is no underlying support either from operating system or software level to predict the user’s intensions based on user context. The main objective of this research is to come up with a framework to predict user intentions based on user context by identifying activity patterns. The framework must be run in-device so that it will function irrespective of the network connectivity. We selected “clustering” as the approach because it does not involve high computation power or complexities to run in-device. We identify activity patterns by clustering the user’s actions and then predict based on the closest cluster for the given time. We have evaluated K-means and Expectation-maximization (EM) clustering algorithms for compatibility for the framework. Unlike computers, mobile devices do not have powerful CPUs or memory. Therefore, we measured CPU time and memory usage of these algorithms to select the best. To maintain low-end device compatibility, we tuned in the algorithm parameters to achieve high accuracy keeping the CPU and memory consumption in low levels. In conclusion, we have successfully identified that EM clustering is suitable for high-end devices and it gives high accuracy while K-means is suitable for low-end devices with acceptable accuracy. We have implemented the framework as an Android library and developed a proof of concept application by embedding the implemented library to show that this research will actually enables application developers to give better user experience to their applications.