Search

To cope with rapid increases in both the number of smart devices and the amount of computation-intensive edge applications on such devices, distributed cross-device edge computing systems have received much attention in the Internet-of-Things (IoT) community. The heterogeneity of IoT devices and the dynamic nature of edge computing environments make it hard to manage the edge resources in a uniform manner that enables collaboration and sharing of resources among the devices. In this paper, we propose Intelligent Distributed Edge Computing (IDEC) system architecture, which is a full-stack system design to support deep learning (DL) model training and inference across heterogeneous resource-constrained IoT devices. IDEC has the characteristics of heterogeneity-compatibility, high performance, and intelligent adaptability. It mainly comprises of three modules: edge resource management, computing task decomposition, and Intelligent Computing Task Allocation (ICTA). IDEC can conduct unified resource management and effective scheduling of heterogeneous edge devices, operator-level task decomposition of deep models, and automatic end-to-end optimization of task allocation based on DL algorithm, thereby achieving resource utilization maximization and further obtaining better system performance. Moreover, through training ICTA periodically in an end-to-end manner based on a continuous learning mechanism, IDEC will become smarter while being used, and help realizing intelligent edge applications and services.

Vehicle's high mobility and the limited coverage of edge servers may result in interruption of ongoing edge services. Therefore, how to ensure the high reliability and availability of services needs to be solved. In this paper, we propose a service continuity management method that considers vehicle behavior, energy consumption, and application redirection in the fifth generation (5G) mobile networks. Firstly, we predict vehicle destination by adopting extreme gradient boosting (XGBoost) to decide where services are migrated following vehicle mobility. Secondly, we propose a MEC handover strategy based on markov decision process to reduce the overall energy consumption. Thirdly, we illustrate application redirection process in application domain based on 5G core network. To evaluate our method, we conducted experiments using open source dataset and simulation data. Our experimental results show that the proposed service continuity management method can achieve a destination prediction accuracy of 87.77%, and can significantly reduce overall energy consumption by 38.7% as compared to frequent mlaration strategy.

With the growing popularity of Internet-of-Things (IoT), large amount of IoT devices are widely used in many spheres of life and tremendous volume of data (which are real-time, streaming and high-dimensional) have been collected, exchanged and processed. In order to reduce the computational complexity and communication cost, it is necessary to reduce the high dimension of these collected data. Because of the data complexity and limited capacity of IoT devices, most of dimensionality reduction algorithms are no longer convenient. To overcome this problem, we propose a lightweight Instance-based Neighbor Embedding (INE) method to accomplish the dimensionality reduction for the IoT streaming data on the resource-constrained devices, which not only keeps the results stable but also maintains the similarity structures of data. Besides, we also propose some techniques and detailed proofs for efficient computation. Extensive experimental results validate the effectiveness of our method.

Internet-of-Things (IoT) plays a critical role in many intelligent scenarios, such as home automation, health care, connected vehicles and so on. Taking into account of several practical concerns like privacy, latency and battery issues, embedding machine learning algorithms in IoT devices is a potentially effective solution. Due to the fact that IoT devices are deployed in highly dynamic environment, reconstruction capability for model which can make the devices adapt to changes of the environment is imperative. In this paper, we design a novel autonomous model training mechanism based on curriculum learning which can deduce a set of labels to rebuild a new model adapting to dynamic environment. In particular, the label learning adopts a three-step strategy, including (1) curriculum generation, (2) curriculum refinement, and (3) curriculum teaching. In the first step, we design a reliability evaluation mechanism to pick out a high-quality set with higher reliability over the original curriculum. In the second step, based on confusion matrix and label propagation, we adjust the labels of the curriculum to accomplish the curriculum refinement. In the third step, the labels of the low-quality set will be adjusted based on the refined curriculum. Then an accurate training set can be obtained, based on which a new model can be built. Extensive experiments on three datasets demonstrate the efficiency of our approach compared to the state-of-the-art methods.

With the emergence of the Internet-of-Things (IoT), we are witnessing rapid increases in the use of wearable devices, which are typically used to monitor human activities and may provide, by leveraging machine learning techniques, underlying contextual characteristics of the human wearers. Upon changes in the IoT environment, most machine learning algorithms need to be retrained-a process that needs large amount of labeled training data. In this paper, we focus on the situation where a new wearable device is added to an existing IoT system which has finite computation ability and memory space, and investigate how to improve the recognition accuracy with the newly added wearable device-without labeling any new instances. We propose a novel algorithm referred to as Domain Adaptation Transduction (DAT), for autonomously reconfiguring the wearable algorithms. Our experimental results on commonly-used benchmark datasets show that the proposed DAT algorithm yields better performance than existing methods when the labeled data is small.

With the emergence of Internet-of-Things (IoT), we are witnessing rapid increases in the amount of sensory data, which are streaming, volatile, often real-time, and heterogeneous in nature, when we enable interactions with the physical world by using sensors and intelligent objects. To help derive value-add insights into the IoT systems and their users, we need methods to automatically extract machine-readable concepts from raw sensory data. To date, only limited research effort has been devoted to automatically extracting (machine-readable) concepts from large sets of heterogeneous multivariate time-series sensory data. In this paper, we propose a framework for real-time automatic concept-extraction in the IoT environments. We enhance the symbolic aggregate approximation (SAX) algorithm into an optimized version (referred to as MultiSAX) for multivariate time-series sensory data - thereby transforming the multivariate time-series sensory data into symbolic representations. By redefining the distance function and density function of the symbolic representations, we extend clustering algorithm and automatically group symbolic representations into different concepts, which can be further operated on by utilizing a rule-based mechanism to make the concepts machine understandable. Our evaluation results show that our proposed method, operating on heterogeneous multivariate sensory data, can achieve automatic concept-extraction with low construction error and low communications overhead.

As the Internet of Things (IoT) expands its reach into virtually every domain, high-speed data processing and shorter response time are becoming more necessary than ever. Embedding machine learning algorithms in IoT devices is a possible effective solution. However, the IoT devices are typically deployed in highly dynamic and uncontrolled environments, which makes machine learning algorithms with automatic learning mechanisms imperative. In this paper, we introduce a framework for synchronous automatic learning. By defining the concept of the centroid distance vector we propose synchronous automatic learning framework which automatically accomplish the reconfiguration of machine learning algorithm based on label refinement. Our evaluation shows that the automatic reconfiguration of machine learning algorithms is achievable with higher accuracy compared to the state-of-the-art methods.

Wearable devices play an essential role in the Internet-of-things applications. By leveraging machine learning algorithms, they can achieve activity recognition, which is an essential function for Internet-of-Things applications, such as home automation, intelligent health care, connected vehicles and so on. However, because of the high substitutability, the wearable devices always have short lifetime, which makes the wearable devices with automatic model construction ability imperative. In this paper, we introduce an automatic model construction method for newly added wearable device, which can automatically construct the machine learning model without the need for any newly labeled training data. The method adopts the self-paced learning regime, which ensures the quality of the learning curriculum and further improves the accuracy of the automatically reconstructed model. We conduct empirical evaluation of our method compared with four state-of-the-art methods; and the evaluation shows that by applying our automatic construction method, we can achieve the reconstructed machine learning model with higher accuracy and lower computational cost.

Fuel consumption is an important factor in the route search for vehicular navigations. In this paper, we propose a personalized eco-friendly route search method that considers the driver's driving style, road traffic, geographic information and vehicular parameters. Firstly, we classify the driving styles into three categories (calm, normal and aggressive) by adopting a Learning Vector Quantization (LVQ) neural network with inputs based on 16 characteristics related to vehicle speed, acceleration and engine speed. Secondly, we design a roadway traffic estimation model based on functional similarity and congestion propagation characteristics. Thirdly, we propose a model for fuel consumption estimation (FCE) based on multivariate nonlinear regression to accomplish the eco-friendly route search. To evaluate our route search method, we conducted experiments using real-world vehicle data gathered in the city of Beijing. Our experimental results show that the proposed route search method can achieve a driving-style prediction accuracy of 82.17%, and can reduce the fuel consumption by 16% as compared to the time-priority routes.

The assessment of node importance has been a fundamental issue in the research of complex networks, which is of overreaching importance to improve the robustness of actual system and design efficient system structure. Most researchers use the betweenness centrality (BC) or Katz centrality (KC) to measure node importance. However, the betweenness only takes into account the shortest path, regardless of the non-shortest path. The Katz centrality gives different weights to all the paths in the network, but the ranking result is close to the result of local path index. Therefore, a new algorithm combines betweenness centrality and Katz centrality(BKC) is proposed, which considers both the local node importance and the global node importance comprehensively, and overcomes the limitations that the node importance evaluation only depending on the adjacent nodes. Experimental results on the kite network and the APRP network illustrate that BKC is more effective to find out important nodes in different types of complex networks than the algorithms compared. Moreover, the cascading failures on the author collaboration network also illustrate that the networks are more vulnerable when continuously removing the important nodes identified by BKC, which further proves the effectiveness of our method.

As understanding social networks requires complete and real-time collection of social network data, it is critical to develop efficient data gathering methods. In particular, smart deployment of monitoring points on these networks will greatly help the data acquisition process. In this paper, we propose a monitoring point deployment strategy on social network data based on the set cover problem of user relationships, and design a greedy solution. We further use data sets derived from real social networking sites (Twitter and Sina Weibo) and a synthetic scale-free network to evaluate the proposed solution in term of the monitor capacity and the required number of monitors. The results show that the proposed solution can achieve good coverage and timely access to social network data.

With the rapid development of Internet, we have to face an increasingly important reality regarding the way in which people propagate information and express their thoughts. Meanwhile, social network has caused the tremendous changes of information propagation model. On this premise, perceiving information propagation model effectively is critical in understanding the rules of information propagation. In this paper, we determine the relationship between modularity and information propagation model, and propose a novel method which combines modularity with triple exponential smoothing model to perceive information propagation mechanism. The experimental results demonstrate the effectiveness of our method in information propagation model perceiving. This method can also predict information propagation trends to a certain extent.

Trust between users in social network plays an important role on information diffusion, rumor propagation and social recommendations. Therefore, considering the universality and negative influence of rumors, it is crucial to look into the trust in detail among social network users when try to design effective immunization strategies for rumor propagation. In this study, we build a weighted trust network, in which the strength of tie refers to trust among users. Based on the weighted trust network, we further propose an immunization strategy (called MST Immunization) concerning maximum spanning tree. By simulation and empirical experiments, we compare the MST immunization strategy with random strategy and targeted strategy. The results show that this new immunization strategy achieves high immunization performance and better maintaining of network connectivity.

Using the complex network methods, we have constructed bipartite networks of investment relationship between CSl 300 constituent stocks and their top 10 investment fund shareholders with data released on December 31, 2010. Here we analyze a variety of structure properties of these networks, including the average distance, clustering coefficient, and overlapping community structure. Studying the structure of this graph is essential for understanding the structural properties of the investment relationship it represents. The scale-free and small world properties of the funds network and the stock network are found, which gives a way to explain the investment policy adopted by institutional investors. The community structure of the fund network shows that few fund investors have great influence and certain investment funds in the same community have convergence in investment philosophy.

Deviation for a practical electro-hydraulic deviation control system, mathematical model is set up. Fuzzy control is introduced on the basis of PID control and FUZZY-PID dual-mode fuzzy controller is designed. PID and fuzzy control can be switch by setting the error value. Simulation results show that the system retains the respective advantages of the two controllers, and control performance is greatly improved.