A Study On Context Recognition And Mining Of Mobile User Data

Recent years have witnessed a revolution in Mobile Internet. Mobile device has become a new way for user to use the Internet. However, Internet service providers do not understand users, thus cannot provide personalized and context-aware services. Mining the context data of mobile user is a novel approach to understanding user, which can provide a solid basis of personalized and context aware services for web applications. Here context refers to environment informa-tion and user behavior data that the sensors of mobile device can detect. In this thesis, we conduct systematically study of recognizing and mining contexts from mobile user data, and thus helps to understand mobile users semantically.1. We propose an unsupervised approach to modeling personalized context of mobile users. For the problem of user context modeling, it is hard to apply supervised approach since the lack of user labeling data. Thus, we propose to use unsupervised learning model to mining user context. Our approach consists of two steps, in the data preprocessing step, we use a minimum entropy method to segment the user context log to context sessions, and in the modeling step, we use clustering and topic model to learn user context respectively. We use K-means algorithm to cluster context sessions in the space of context feature-value pairs, and extract user contexts from the clusters. To meet the structure constraint of context data, we extend two popularly used topic model MU and LDA to learn contexts from the context sessions. The experiment results on the real user context data show our approach is effective.2. We propose an approach to mining user significant places from Cell ID trajectory data. The significant places are the most important context of users. Most related works focus on mining significant places from GPS data, however as the limited power of mobile device, it is hard to collect enough data using the GPS sensor. Thus, we propose to utilize the Cell ID data. Different from other related work that mining significant places from Cell ID data, we use more additional information such as the real location and coverage area of cell sites. Our approach consists of two steps. In the online step, we detect whether user is in stationary states, if so we compute the stay area of user, and update the significance of grids. In the offline step, given the user request of query for significant places, we use a recursive method to extract significant places with the representation of grids. To evaluate this approach, we have developed a demo system and the experimental results on real trajectory data show that our approach can outperform baseline methods both in precision and recall.3. We propose to integrate context recognition into time management software. Since users of time management are often need to record common contexts, some of them may be tired of such repeated behaviors, and give up using this software. We propose a semi-supervised approach to recognizing user context automatically, thus can help user to record common contexts. We utilize the context records in the time management software, and model user context status and context data with an extended bayesian hidden markov model(HMM). To determine the number of contexts and speed up the training time of HMM, we also propose an extended topic model DP-MUC. The experimental results show that our approach is both effective and efficient.