A Study Of Mobile Environment Oriented Efficient Context Data Mining And Energy Efficient Sensing

To enable the intelligent mobile applications of the future, it is important to understand mobile users through the data collected from their mobile devices. In recent years, more and more commercial mobile devices such as smart phones and personal digital assistants are equipped with multiple context sensors including optical sensors,3D accelerometers, GPS sensors, etc, which makes it possible to bring to bear intelligent context-aware applications to ordinary mobile users. The data collected by sensors reflects user context because the mobile device is often carried by mobile user. Moreover, many interesting knowledge can be discovered from the collected context data (e.g., GPS trajectories and usage log) through data mining technologies. Some mobile applications have been used the data to support more intelligent, personalized service for users. For example, auto-adapting user interface according to light intensity can protects use eyes and extends the battery life. Since mobile internet is the one of the main trends in IT development, mobile context sensing and user behavior mining become more and more attentive in future.There are some new characteristics for mobile context sensing. Firstly, com-puting resource, storage capacity is limited in mobile device, and intelligent ap-plications require high real-time condition. Secondly, compared with traditional context data, mobile context data contains richer context information. It records where user has arrived and some features are sparse(e.g. user interactions), so the traditional data mining technologies can not be directly used in mobile context sensing. In this thesis, we consider these characteristics, and leverage machine learning technologies for context sensing to in-depth study of energy efficient con-text sensing and efficient user behavior mining. To be specific, energy efficient context sensing can be a middleware which provides information to applications’ continues context sensing, and user behavior mining can be used for research in mobile user recommendation or user interest. The main research innovations of the thesis are as follows:Firstly, we propose an efficient algorithm, named BP-Growth, for efficient behavior pattern mining. The existing approaches for mining these behavior pat-terns are not practical in mobile environments due to limited computing resources on mobile devices. To fulfill this crucial void, we investigate optimizing strategies which can be used for improving the efficiency of behavior pattern mining in terms of computing and memory needs. Specifically, we examine typical optimizing s-trategies for association rule mining and study the feasibility of applying them to behavior pattern mining. We use real context data collected from10mobile user for experiments and the experimental results show that BP-Growth outperforms benchmark methods with a significant margin in terms of both computing and memory cost.Secondly, we propose the Current Status Inference (CSI) model for energy efficient context sensing. The battery capacity of mobile devices becomes the bottleneck of context-aware applications because some context sensors are very energy consuming and cannot continuously work for the sake of user experience. We argue that the outputs of different context sensors of a mobile device may be more or less correlated since they essentially capture the same context at each time point, even though from different perspectives. Intuitively, we may be able to selectively avoid invoking high energy consuming sensors by inferring their statuses from the outputs of other sensors. Thus we group context sensors into the basic sensors, the light-duty sensors and the heavy-duty sensors based on their energy consumption and function, and then propose a CSI model. A CSI model will infer the status of a heavy-duty sensor according to output of basic/light-duty sensor. If the heavy-duty sensor is in stable status, the system will not invoke it again and instead use the latest invoked value. The experimental results on real data sets show that the energy efficiency of GPS sensing and audio level sensing are significantly improved by the proposed approach while the sensing accuracy is over90%.Finally, we propose the Status Interval Inference (SII) Model for energy effi-cient context sensing. In real world, a mobile user is generally in a state continuing for a period of time. A well trained model can be used to infer the time how long the heavy-duty sensor will stay in stable status. To be specific, we propose a SII model which can use the output of basic/light-duty sensor to infer the interval time how long the heavy-duty sensor will stay in stable status. If the heavy-duty sensor will switch to unstable status after some intervals, the system will not invoke it until these intervals are over and instead use the latest invoked value. Then we summary the CSI and SII models to a general Stable Status Inference (SSI) model and propose its framework and use maximum update interval for reducing the probable accumulated errors in their estimations. The experimental results also show that the energy efficiency of GPS sensing and audio level sensing are significantly improved by the proposed approach while the sensing accuracy is high and stable.