Dianxi Shi

A novel orientation- and location-independent activity recognition method

The orientation and location of a mobile phone pose fundamental challenges to activity recognition (AR) in a device. Given that AR significantly affects recognition accuracy, in this study, we focus on eliminating the influence of orientation and location changes on AR. First, we propose an activity recognition framework, which is independent of orientation and location changes, to uniformly deal with the problem of orientation and location changes on AR. Second, a dynamic coordinate transformation approach on inertial sensor data is proposed. In this method, the data collected in different orientations are dynamically mapped to the reference coordinate system of a mobile phone. The classification on the mapped data can reach significantly higher accuracy than that on the original data. We design four sets of comparative experiments to verify the validity of the proposed method, and the results demonstrate its effectiveness. Third, the influence of the location changes of mobile phones on AR is eliminated through the location-specific AR method. The effectiveness of the proposed method is verified by two groups of contrast tests. Finally, a real-time AR system is implemented on an Android platform. Results demonstrate that the proposed method obtains valid recognition results despite various orientation and location changes.

SASLL: A System Annotating Semantic Label of Location

Nowadays GPS embedded in mobile device such as smartphones can easily identify peoples physical locations. However, in daily life people are more concerned about semantic locations (such as dormitories, laboratories, shopping malls, etc.). Usually GPS positioning uses continuous sampling method, which results in a lot of semantically independent sample points. We call these points outliers. How to remove outliers from GPS data and thereby cluster meaningful semantic places is a research challenge in current field of pervasive computing. Aiming at the characteristics of this problem, we first propose a novel approach to add semantic annotations to newly discovered places every day. We use an unsupervised method to discover semantic places, which ensures accuracy of the results and reduces the amount of calculation. Secondly, we discuss the concept of outliers in GPS data collected in daily life, and then eliminate outliers using a density-based method. Moreover, we perform experiments to validate its effectiveness. Thirdly by taking advantage of rule-based inference and reverse geocoding we proposed an approach to calculate the probable semantic labels, which can help user annotate places and reduce the burden on users. Finally, we develop a local System Annotating Semantic Label of Location(SASLL) and by carrying out experiments we demonstrate the validity of our research.

Data-intensive service system: Model and analysis

Cloud computing has been attracting more and more enterprises and researchers, and one goal of it is that computing is provided on demand as a service like water and electricity. Meanwhile, data-intensive service like MapReduce jobs is widely used in data centers. We model the data-intensive service system as an M/M(t)/n/k queueing system, where the input process can be seen as Poisson process, the mean service time is a function of time, and the system resources can be divided into n servers and a queue of maximum length of k. We analyze the simplified system where service provider enlarges the system scale in order to provide constant time service. We also study the problem of how to maximize profit for service provider.

Activity Recognition Based on the Dynamic Coordinate Transformation of Inertial Sensor Data

The orientation of a mobile phone is a fundamental challenge in activity recognition (AR) on this device that significantly affects recognition accuracy. In this study, we propose a novel orientation-independent method to eliminate the influence of orientation changes on AR. This method is a dynamic coordinate transformation approach on inertial sensor data. In this method, the data collected in different orientations are dynamically mapped to the reference coordinate system of a mobile phone. The classification on the mapped data can reach much higher accuracy than on the original data. We evaluated our method using four sets of comparative experiments. The proposed method was compared with initial processes without considering orientation, two other methods in previous studies. The experiments showed that the proposed method outperformed the other three approaches. Implementing a real-time activity recognition system on an Android platform demonstrated that the proposed method obtained valid recognition results despite various orientation changes.

A Framework of Fine-Grained Mobile Sensing Data Collection and Behavior Analysis in an Energy-Configurable Way

In recent years, mobile sensing data are widely used for analyzing humans activities, usage patterns, emotions, health conditions and social relationships. In order to understand and analyze humans behaviors, several frameworks have been proposed to collect mobile sensing data. In this paper we extend previous works and design StarLog, which is a distributed and energy-configurable framework for both mobile data collecting and analyzing. It collects fine-grained sensing data of five categories, reflecting users locations, activities, interactions with smart phone, social contacts and device setting habits. Data analyses are developed on both client side and server side to understand individual as well as crowd behaviors. Besides, StarLog proposes optional modes for collecting sensory data from GPS, accelerometer, gyroscope and magnetometer to make it configurable for battery concern. To achieve the data, we recruited 21 students and processed the data collection for 20 days. Up to 17 kinds of data sources were collected, including data sources like micro messages (Weixin and QQ) that have never been collected before. Analyses of these data leaded to interesting findings in users patterns and social ties.

Random Walk Based Inverse Influence Research in Online Social Networks

In online social networks, social influence of a user reflects his or her reputation or importance in the whole network or to a personalized user. Social influence analysis can be used in many real applications, such as link prediction, friend recommendation and personalized searching. Personalized Page Rank, which ranks nodes according to the probabilities that a random walk starting from a personalized node stops at all nodes, is one of the most popular metrics for influence analysis. In this paper, we study the problem of inverse influence in online social networks. Different from Personalized Page Rank, the inverse influence for a personalized node ranks nodes according to the probabilities that all nodes stop at the personalized node in limited steps. We propose two computation models for inverse influence, i.e., the random walk based and the path based. Both of the models have high computation complexity, and cannot be used in large graphs, so we propose a Monte Carlo based approximation algorithm. Experiments from synthetic and real world datasets show that, our algorithm has equivalent or even better accuracy than related researches in link prediction, and thus can be used in friend recommendation in online social networks.

MCHITS: Monte Carlo based Method for Hyperlink Induced Topic Search on Networks

Hyperlink Induced Topic Search (HITS) is the most authoritative and most widely used personalized ranking algorithm on networks. The HITS algorithm ranks nodes on networks according to power iteration, and has high complexity of computation. This paper models the HITS algorithm with the Monte Carlo method, and proposes Monte Carlo based algorithms for the HITS computation. Theoretical analysis and experiments show that the Monte Carlo based approximate computing of the HITS ranking reduces computing resources a lot while keeping higher accuracy, and is significantly better than related works

Random Walk Based Location Prediction in Wireless Sensor Networks

With the development of wireless sensor network (WSN) technologies, WSNs have been applied in many areas. In all WSN technologies, localization is a crucial problem. Traditional localization approaches in WSNs mainly focus on calculating the current location of sensor nodes or mobile objects. In this paper, we study the problem of future location prediction in WSNs. We assume the location histories of mobile objects as a rating matrix and then use a random walk based social recommender algorithm to predict the future locations of mobile objects. Experiments show that the proposed algorithm has better prediction accuracy and can solve the rating matrix sparsity problem more effectively than related works.