Multivariate Time Series (MTS) forecasting is crucial in many domains, such as financial market analysis, weather forecasting and energy management. Among various solutions for this task, transformer-based models are popular due to their excellent ability capturing long-term dependencies. However, those models still face challenges for precise MTS forecasting due to the ignorance of variables' correlations and the complexity of the attention calculation. To address these challenges, we propose a novel Variable Adaptive Embedding transformer, namely VAEformer. Specifically, on the one hand, we employ variable adaptive embedding to capture correlations among variables better. On the other hand, we employ low-complexity linear layers to model temporal dependencies to improve the efficiency. Furthermore, the excellent predictive performance on real-world datasets from various domains suggests that our VAEformer can serve as a cross-domain tool, fostering collaboration and coordination among researchers from different fields.

The attention mechanism has the advantage of handling long-term correlations, and has been widely adopted in multivariate time series (MTS) prediction.As an important application of MTS, traffic flow prediction has the most popular solution using transformerbased prediction models nowadays.Just with attention mechanism, those models can learn the spatio-temporal correlations from traffic data.However, the up-to-date linear prediction models have questioned the effectiveness of current transformer-based models in certain conditions, which provides new possibilities for more efficient work.We rethink the role of the attention mechanism during spatio-temporal modeling from a decoupling perspective, and propose DEC-Former for traffic flow prediction.Specifically, the trend and seasonal parts of the time series data, the geographical adjacency of the nodes in the road network, and the traditional encoderdecoder architecture, are respectively decoupled.Such decoupling leverages the attention mechanism's advantage to capture longterm and long-range correlations.From extensive experiments on four real-world datasets, our work proves better predictive performance and efficiency than state-of-the-art attention-based models.Two case studies further show the distinct real effects.

As the development of cities, traffic congestion becomes an increasingly pressing issue, and traffic prediction is a classic method to relieve that issue. Traffic prediction is one specific application of spatio-temporal prediction learning, like taxi scheduling, weather prediction, and ship trajectory prediction. Against these problems, classical spatio-temporal prediction learning methods including deep learning, require large amounts of training data. In reality, some newly developed cities with insufficient sensors would not hold that assumption, and the data scarcity makes predictive performance worse. In such situation, the learning method on insufficient data is known as few-shot learning (FSL), and the FSL of traffic prediction remains challenges. On the one hand, graph structures’ irregularity and dynamic nature of graphs cannot hold the performance of spatio-temporal learning method. On the other hand, conventional domain adaptation methods cannot work well on insufficient training data, when transferring knowledge from different domains to the intended target domain.To address these challenges, we propose a novel spatio-temporal domain adaptation (STDA) method that learns transferable spatio-temporal meta-knowledge from data-sufficient cities in an adversarial manner. This learned meta-knowledge can improve the prediction performance of data-scarce cities. Specifically, we train the STDA model using a Model-Agnostic Meta-Learning (MAML) based episode learning process, which is a model-agnostic meta-learning framework that enables the model to solve new learning tasks using only a small number of training samples. We conduct numerous experiments on four traffic prediction datasets, and our results show that the prediction performance of our model has improved by 7% compared to baseline models on the two metrics of MAE and RMSE.

Abstract Workflow management system (WfMS) in cloud always works as platform as a service to manage customized business processes for massive enterprises. In big data era, non‐functional guarantees of such systems are significant when facing a large number of users and concurrent requests. It is not trivial to support multi‐tenancy and hold high‐availability, because traditional architecture cannot simultaneously satisfy requirements about data isolation and runtime efficiency. In this paper, a modularized distributed workflow management system is proposed, which considers both multi‐tenancy and high‐availability in storage and engine parts of the system. A multiple‐worker‐with‐separate‐schema mechanism is defined to jointly manage the data for tenants, and a proactive strategy is presented to intelligently dispatch large concurrent requests from users to engine workers. After extensive case studies and experiments in practical scenes, our system deployed on modest machines is proved to support tens of thousands of tenants, second‐level response time for 10 K concurrency, and no‐human‐intervened failure recovery for a fail‐stop system node.

Information extraction and its service is a hot topic. Many works focus on extracting information from a certain web page and ignore the localization of the webpage which contains useful information. Nevertheless, developing a holistic system to extract information consists of locating a webpage and extracting information from that webpage, and these two steps are indispensable. For instance, extracting lecture news from universities' websites is a typical hard task that need to locate web pages and extract news information from them. Due to different layouts and visual appearances, statistic-based methods and visual based methods failed to find them. In this study, we propose an all-holistic method to locate lecture news on the university website. Graph Convolutional Network (GCN) is applied to fuse the multimodal data, which could learn useful features from different views, the linked relationship, the visual similarity, and the semantic of web pages. Firstly, we apply the link model to explore the parent-child relationship between web pages, then calculate the similarity of parent-child pages using a visual model and obtain the semantic features based on the BERT model. Specifically, the visual similarity features are learned based on triplet loss function which imposes the Convolutional Neural Network (CNN) model to learn similar parts in the same group. Lastly, these features are fused into the GCN model to find a certain webpage and it can be adaptive to various university websites. The experiments conducted on 50 websites show our method outperforms state-of-the-art.

A Workflow management system (WfMS) is usually deployed in a Cloud environment in a distributed architecture, and the execution of the process task requires the workflow engine to parse the process definition in its cache. However, the current WfMS seldom considers the problem of the cache hit, which leads to repeated parse and decreases request response time and service quality of WfMS. A task execution dispatching method is proposed in this paper based on cache mechanism. The method enables each task’s requests to be dispatched to the best matching worker, and comprehensively utilizes the cache in the worker or the database. After experiments of practical scenes, compared with two traditional ones commonly used in business, our method is proved to improve the cache hit rate, reduces the response time, and improves system performance.

As a transportation way in people’s daily life, highway has become indispensable and extremely important. Traffic flow prediction is one of the important issues for highway management. Affected by many factors, including temporal, spatial, and other external ones, traffic flow is difficult to accurately predict. In this paper, we propose a graph convolutional method. And the name of our model proposed is the hybrid graph convolutional network (HGCN), which comprehensively considers time, space, weather conditions and date type to achieve better predicted results of traffic flow at highway stations. Compared with baselines implemented by various machine learning models, all metrics of our model are reduced dramatically.

Many anomaly detection algorithms have been provided as services for more convenient and efficient utilization in IoT era. Due to concept drift existed in dynamic IoT stream data, it is a changeling task to apply proper anomaly detection services at run time. For effective on-line anomalous data discovery, this paper proposes a service selection framework to dynamically select and configure anomaly detection services. A fast classification model based on XGBoost is trained to identify the pattern of various stream data, so that suitable service can be selected and configured according to the pattern of stream data. Extensive experiments on real and synthetic data sets show that our framework can select suitable service for different scenarios, and the accuracy of the chosen services outperforms state-of-the-art methods.

Abundant sensors in various types are widely used in modern cities to comprehend the current situations in real time. The raw data in open conditions is always in low quality and is hard to employ directly due to its imperfect or missing records. Traditional data preprocessing methods focus on the offline historical data and remain a dilemma between the efficiency and the overhead. In this paper, a data harmonization service DS-Harmonizer is proposed on spatiotemporal data stream in the edge computing environment. Through the online cleaning and complementing steps of the hierarchical service instances, the records’ validity and continuity can be guaranteed in an efficient way. On the simulated data in a practical project, the service shows high performance, low latency, and acceptable precision in extensive conditions.

In Big Data era, continuous data with low latency and high throughput makes high-availability essential for stream computing. Traditional availability guarantee is tightly-coupled and inefficient for customization and reuse. In this paper, a framework is proposed to improve the availability of stream computing, in which basic functions are provided as general services like reliable point-to-point communication and distributed status management. With its help, high-level patterns can be achieved effectively. Comprehensive experiments have been designed and evaluated to show the availability improvement with acceptable extra overheads.

With the continuous expansion of the scope of traffic sensor networks, traffic sensory data becomes widely available and is continuously being produced. Traffic sensory data gathered by large amounts of sensors show the massive, continuous, streaming and spatio-temporal characteristics compared to traditional traffic data. In order to satisfy the requirements of different applications with these data, we need to have the capability of processing both real-time traffic sensory data in streaming way and historical traffic sensory data in large amount. In this paper, we present an approach and corresponding system for traffic sensory data processing, which is designed to combine spatio-temporal data partition, parallel pipeline processing and stream computing to support traffic sensory data processing in a scalable architecture with real-time guarantee. Three types of applications in real project are also described in detail to show the significant effect gains of the proposed approach and system. Numerical evaluations according to experiment results also show that the system can gain high performance in terms of the processing time of traffic sensory data stream.

With the continuous expansion of the scope of traffic sensor networks, traffic sensor data becomes widely available and large in amount. Traffic sensor data gathered by large amounts of sensors shows the massive, continuous, streaming and spatio-temporal characteristics compared to traditional traffic data. In order to satisfy the requirements of different applications in Intelligent Transportation Systems (ITS), we need to have the capability of real-time processing over both streaming and historical traffic sensor data. In this paper, we present DeCloud4SD, an integrated processing platform for traffic sensor data, which is designed to provide services for receiving, storing, acquiring and computing traffic sensor data in a scalable architecture with real-time guarantee. Three types of applications using DeCloud4SD in a real ITS project are also described in detail. Through the analysis of these applications, we can see that DeCloud4SD can ensure: 1) scalable and customizable traffic sensor data gathering and computing, 2) rapid application development and deployment using a MapReduce-like model, 3) seamless integration with existing relational data sources and applications.

SUMMARY Under distributed Cloud environment, the real‐time and continuous data stream makes the availability during processing essential but expensive. For aggregation tasks of data stream processing systems, traditional replica‐based high‐availability mechanisms require large overheads at run‐time and long recovery latency at fail‐time, because of specific nature of aggregations. In this paper, we focus on the typical quantile tasks and propose a feature‐based high‐availability mechanism to reduce related overhead and the latency. With the help of monitor module, quantile feature is maintained incrementally through histogram synopsis over time‐based sliding window, and the failed quantile tasks can be recovered precisely with high probability in an efficient way. The effectiveness has been analyzed theoretically, and meanwhile, the acceptable tradeoff between overheads and performance has been demonstrated by comprehensive experiments on both synthetic and real data. Copyright © 2013 John Wiley & Sons, Ltd.

Data process in Cloud or IoT (Internet of Things) sometimes implies continuous real-time queries as data streams. In order to acquire extreme value of data stream over time-based sliding window, traditional approaches computed the exact solution through vast space especially under ultra circumstances like high-rate or high-concurrency. In this paper, we design space-bounded synopsis data structure and extreme aggregation algorithm to get approximate solution by finite extreme candidates over time sliding window, whose validity can be theoretically guaranteed. Comprehensive experiments over synthetic and real data set are designed to analyze the tradeoff between accuracy and overhead, which also illustrate the efficiency.

Web services are rapidly emerging as a popular standard technology for sharing data and functionality among heterogeneous systems. Service providers and consumers are loosely coupled and distributed across the network, either within an organization or across organizational boundaries, and therefore, performance becomes a major concern in such a distributed environment. Furthermore, XML is widely used as message format for service providers and consumers in Web services environment. XML message packaging and parsing brings extra overhead to both ends. Web services response latency, as well as throughput, is becoming a bottleneck problem. In this paper, We propose a consistency-preserving mechanism for Web services response caching, which reduces the volume of data transmitted without semantic interpretation of service requests or responses, and accelerates the services response finally. It achieves this reduction through the use of cryptographic hashing to detect similarities with previous results. Experiments with an initial prototype called SigsitAcclerator indicate that our mechanism can lead to significant performance improvement over more straightforward techniques.