Improving Mobile Services Through Multi Context Analysis PDF Download

The development of ubiquitous computing and context-aware services has profoundly changed the way users interact with their devices. The increasing sensing and computational power possessed by state-of-art mobile devices promotes context-awareness to the next level where a full range of contexts are ready for exploitation. In contrast to previous research efforts that push the limit for handling of individual context, we believe the time is ripe for incorporating multi-context analysis into context-aware services. Multi-context analysis is a decision making process based on analysis of information obtained from multiple sensing channels and context dimensions. The centerpiece of the analysis is to create a model that integrates available cross-domain knowledge about interesting contexts, discover relationships among them, and determine a relevant set of the knowledge for specific applications. The diversity of available sensors and contexts that enables multi-context analysis at the same time poses several challenges on conducting such analysis. We start with a number of case studies where specific applications are defined and cross-domain knowledge required are easily identifiable. We are particularly interested in one of the most important contexts, location context namely, and location-based mobile services. We first look at the limitation of traditional indoor localization systems with Wi-Fi fingerprinting, and propose a new method that applies multi-context analysis of Wi-Fi and Bluetooth signals to improve the bootstrapping efficiency while retaining the accuracy requirement. Furthermore, indoor navigation system previously relying on Wi-Fi localization system is studied, and improved by exploiting the signal landmarks recognized by multi-context analysis of inertial sensors and wireless signals. Privacy concerns associated with distribution of location contexts are also investigated in this research. We identify the risk of an advanced adversary compromising continuous location privacy with multi-context analysis of real-time location updates and a-priori knowledge, and propose a perturbation scheme to counter such effort. Based on experience and insight gained from the case studies, we explore modeling techniques that can facilitate the extraction and encoding of cross-domain knowledge needed for multi-context analysis, preferably through an automatic process. A Bayesian graphical model is proposed to represent contexts and encode their statistical correlation in describing different situations. Such a model is evaluated with a context attestation and validation system that detects falsified user context claims. Given the increasingly diverse and heterogeneous cross-domain knowledge about contexts available to us, it is more and more difficult for human to recognize which piece of information is indeed relevant. To address this, a deep learning framework is explored to enable the abstraction and fusion of cross-domain knowledge from raw sensing data. Research reported in this thesis covers the general life-cycle of contexts from acquisition, distribution, verification, to final consumption. Results from these studies confirm that multi-context analysis can greatly benefit performance and user experience in context-aware services. As part of the future work, we plan to extend the modeling technique to cross-domain knowledge of extended types and granularity so eventually a general framework can be developed to improve majority of context-aware services through multi-context analysis.