With the proliferation of cloud services and the continuous growth in enterprises’ demand for dynamic multi-dimensional resources, the implementation of effective strategy for time-varying workload scheduling has become increasingly significant. In this paper, we propose a deep reinforcement learning (DRL)-based method for time-varying workload scheduling, aiming to allocate resources efficiently across servers in the cluster. Specifically, we integrate a classifier and queue scorer to construct a priority queue that exploits temporal resource utilization patterns across different workload classes. Then, we design parallel graph attention layers to capture the dimensional features and temporal dynamics of cloud server cluster. Moreover, we propose a DRL algorithm to generate scheduling strategies that can adapt to dynamic environments. Validation on real-world traces from Google cluster demonstrates that our method outperforms existing approaches in key metrics of cloud server cluster management.

Contrastive learning (CL) has been widely used to improve recommendation performance, since its self-supervised signals can effectively alleviate the data sparsity issue in recommender systems. Nevertheless, most existing CL-based recommendation models construct negative sample pairs following the common practice, which may include many false negatives (i.e., highly similar nodes). Some studies detect potential false negatives through specific rules and either remove these samples from the negative samples or treat them all as positive samples. However, they lack a finer-grained consideration of the samples. To address this limitation, we propose a Ranking Enhanced Fine-grained Contrastive Learning method (REFCL). Specifically, we devise two sampling strategies to generate two sets of positive samples with different confidence levels. These well-ordered positive sample sets are then integrated into our novel contrastive loss, which allows for a more nuanced consideration of distinctions between samples and thereby enhancing the discriminative qualities of the learned user/item representations. Extensive experiments conducted on three real-world datasets demonstrate the effectiveness and generality of REFCL.

Multivariate Time Series (MTS) forecasting is to accurately predict future trends through in-depth analysis of historical time series data, and to provide valuable reference to support for decision-making. Because of its complex time variations, some deep learning methods are applied to model time series, where Graph Neural Networks (GNNs) are included as their powerful capability to model variables’ relationship. However, these methods ignored to either construct the graph structure by considering the dual features of periodic and aperiodic characteristics in time series or dynamically capture the time variations patterns by short- and long-term dependencies. In this paper, we propose a model named Time-Decay Dynamic Graph Neural Network (TDDGNN), which combines period-aware mechanism with a dual-attention time-decay dynamic graph structure. To consider the dual features in time series, Feature Extraction (FE) module is designed to mine the periodic features and aperiodic features. Otherwise, Time-Decay Dual attetion dynamic graph (TDD) module is designed for exploring both short- and long-time intervals. The experimental results show that the TDDGNN model outperforms baseline models on the three datasets. The results of this study will provide a valuable reference scheme for solving the problem of periodic characteristics and correlationship between variables in MTS forecasting.

The development of deep learning and edge computing provides rapid detection capability for surface defects. However, components produced in actual industrial manufacturing environments often have tiny surface defects and training data for each specific defect type is limited. Meanwhile, network resources at the edge of industrial networks are difficult to guarantee. It is challenging to train a proper surface defect detection model for each specific surface defect type and provide a real-time surface defect detection service. To address the challenge, in this paper, we propose a real-time surface defect detection framework based on transfer learning with multi-access edge-cloud computing (MEC) networks. Furthermore, we improve the original YOLO-v5s framework by introducing the spatial and channel attention mechanism, and adding an additional detection head to enhance the detection ability on tiny surface defects. Evaluation results demonstrate that the proposed framework has superior performance in terms of improving detection accuracy and reducing detection delay in the considered MEC network.

Self-supervised learning (SSL) recently has achieved outstanding success on recommendation. By setting up an auxiliary task (either predictive or contrastive), SSL can discover supervisory signals from the raw data without human annotation, which greatly mitigates the problem of sparse user–item interactions. However, most SSL-based recommendation models rely on general-purpose auxiliary tasks, e.g., maximizing correspondence between node representations learned from the original and perturbed interaction graphs, which are explicitly irrelevant to the recommendation task. Accordingly, the rich semantics reflected by social relationships and item categories, which lie in the recommendation data-based heterogeneous graphs, are not fully exploited. To explore recommendation-specific auxiliary tasks, we first quantitatively analyze the heterogeneous interaction data and find a strong positive correlation between the interactions and the number of user–item paths induced by meta-paths. Based on the finding, we design two auxiliary tasks that are tightly coupled with the target task (one is predictive and the other one is contrastive) toward connecting recommendation with the self-supervision signals hiding in the positive correlation. Finally, a model-agnostic dual-auxiliary learning (DUAL) framework, which unifies the SSL and recommendation tasks, is developed. The extensive experiments conducted on three real-world datasets demonstrate that DUAL can significantly improve recommendation, reaching the state-of-the-art performance.

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained sentiment analysis task and has become a significant task with real-world scenario value. The challenge of this task is how to generate an effective text representation and construct an end-to-end model that can simultaneously detect (target, aspect, sentiment) triples from a sentence. Besides, the existing models do not take the heavily unbalanced distribution of labels into account and also do not give enough consideration to long-distance dependence of targets and aspect-sentiment pairs. To overcome these challenges, we propose a novel end-to-end model named Prior-BERT and Multi-Task Learning (PBERT-MTL), which can detect all triples more efficiently. We evaluate our model on SemEval-2015 and SemEval-2016 datasets. Extensive results show the validity of our work in this paper. In addition, our model also achieves higher performance on a series of subtasks of target-aspect-sentiment detection. Code is available at https://github.com/CQUPT-CaiKe/PBERT-MTL.

Network function virtualization (NFV) is a promising paradigm that network functions can be deployed on commodity servers instead of dedicated servers to enhance the resource utilization and reduce the management difficulty. Based on the NFV technology, a complex network service can be composed of a series of ordered virtual network functions, known as service function chain (SFC). In this context, how to efficiently place SFCs in acceptable running time to improve resource utilization and service quality while meeting the constraints of the physical network is a critical issue for infrastructure providers. In this paper, we propose a deep reinforcement learning-based approach called DRL-SFCP for adaptive SFC placement. DRL-SFCP maximizes the long-term average revenue by combining both the graph convolution network which extracts the features of the physical network and sequence-to-sequence model which captures the ordered information of the SFC request to generate placement strategies. It learns to make SFC placement decisions via observations of the corresponding performance of past decisions rather than a hypothetical environment. Extensive experimental results show that our DRL-SFCP can achieve 11.6% and 9.6% improvement in terms of the acceptance ratio and the long-term average revenue, compared with existing benchmarks.

With the tremendous growth of mobile data traffic generated by various devices such as smartphones, smartpads and wearable devices, it is necessary for mobile network operators to introduce revolutionary networking techniques, thereby satisfying service requirements of mobile users. Recently, mobile edge computing (MEC) has been regarded as an effective technique to alleviate the traffic burden on backhaul networks. In this paper, we investigate the issue of mobility-aware content caching and user association for ultra-dense MEC networks by minimizing the system costs. The problem is formulated as a complex pure integer nonlinear programming, which is NP-hard. To address the original long-term optimization problem, we decompose it into a series of one-slot subproblems, and then optimize the short-term subproblem in two phases (i.e., content caching and user association). We further propose a mobility-aware online caching algorithm to achieve content caching, and a lazy re-association algorithm to determine user association based on matching theory. Trace-driven evaluation results demonstrate that the proposed framework has superior performance on reducing system costs.

Traffic flow forecasting has become an increasingly important part of intelligent traffic control and management. This task is challenging due to (1) complex geographic and non-geographic spatial correlations; (2) temporal correlations between time slices; (3) dynamics of semantic high-order correlations along the temporal dimension. To address those difficulties, commonly-used methods apply graph convolutional networks for spatial correlations and recurrent neural networks for temporal dependencies. In this work, we distinguish the two aspects of spatial correlations and propose the two types of spatial graphs, named as geographic graph and semantic hypergraph. Furthermore, we extend the traditional convolution into geographic-temporal graph convolution and semantic-temporal hypergraph convolution to jointly capture geographic-temporal correlations and semantic-temporal correlations. Then we propose a geographic-semantic-temporal hypergraph convolutional network (GST-HCN) that combines our graph convolutions and GRU units hierarchically in a unified end-to-end network. The experiment results on the Caltrans Performance Measurement System (PeMS) dataset show that our proposed model significantly outperforms other popular spatio-temporal deep learning models and suggests the effectiveness to explore geographic-semantic-temporal dependencies on deep learning models for traffic flow prediction.

Because social networks have become a vital part of people's lives, cyberbullying becomes the most common risk encountered by young people on social networking platforms and raised serious concerns in society. Over the past few decades, most existing work on cyberbullying has focused on text analysis. Yet, the cyberbullying develops into multi-objective, multi-channel, and multi-form. Traditional text analysis methods cannot satisfy the diversity of bullying data in social networks. To deal with the new type of cyberbullying, we propose a multi-modal detection framework that takes into multi-modal information(e.g., image, video, comments, time) on social networks. Specifically, we not only extract textual features but also use the hierarchical attention networks to capture the session feature in social networks and encode several media information(e.g., video, image). Based on these features, we model the multi-modal cyberbullying detection framework to solve the new form of cyberbullying. Experimental analysis on two real-world datasets shows that our framework outperforms several existing state-of-the-art models.

Word embedding methods use low-dimensional vectors to represent words in the corpus. Such low-dimensional vectors can capture lexical semantics and greatly improve the cyberbullying detection performance. However, existing word embedding methods have a major limitation in cyberbullying detection task: they cannot represent well on "deliberately obfuscated words", which are used by users to replace bullying words in order to evade detection. These deliberately obfuscated words are often regarded as "rare words" with a little contextual information and are removed during preprocessing. In this paper, we propose a word embedding method called LSHWE to solve this limitation, which is based on an idea that deliberately obfuscated words have a high context similarity with their corresponding bullying words. LSHWE has two steps: firstly, it generates the nearest neighbor matrix according to the co-occurrence matrix and the nearest neighbor list obtained by Locality Sensitive Hashing (LSH); secondly, it uses an LSH-based autoencoder to learn word representations based on these two matrices. Especially, the reconstructed nearest neighbor matrix generated by the LSH-based autoencoder is used to make the representations of deliberately obfuscated words close to their corresponding bullying words. In order to improve the algorithm efficiency, LSHWE uses LSH to generate the nearest neighbor list and the reconstructed nearest neighbor list. Empirical experiments prove the effectiveness of LSHWE in cyberbullying detection, particularly on the "deliberately obfuscated words" problem. Moreover, LSHWE is highly efficient, it can represent tens of thousands of words in a few minutes on a typical single machine.

Recently, mobile edge computing has received widespread attention, which provides computing infrastructure via pushing cloud computing, network control, and storage to the network edges. To improve the resource utilization and Quality of Service, we investigate the issue of task offloading for End-EdgeCloud orchestrated computing in mobile networks. Particularly, we jointly optimize the server selection and resource allocation to minimize the weighted sum of the average cost. A cost minimization problem is formulated underjoint the constraints of cache resource and communication/computation resource of edge servers. The resultant problem is a Mixed-Integer Non-linear Programming, which is NP-hard. To tackle this problem, we decompose it into simpler subproblems for server selection and resource allocation, respectively. We propose a low-complexity hierarchical heuristic approach to achieve server selection, and a Cauchy-Schwards Inequality based closed-form approach to efficiently determine resource allocation. Finally, simulation results demonstrate the superior performance of the proposed scheme on reducing the weighted sum of the average cost in the network.

Online music recommender systems are becoming increasingly prevalent because of the popularity of digital music and music recommendation generally caters to users by discovering songs that match their preferences. However, these systems have to face a challenge: how to recommend new songs in a situation where prior knowledge is scarce. Some researches take auxiliary information into consideration in new recommendation approaches to deal with this problem. Nevertheless, they rarely pay attention to complex relationships among different feature spaces when they map those information to a latent space. To this end, this paper proposes an approach that uses non-linear transformation to integrate different auxiliary information into the songs latent representations. Unlike other studies which directly map auxiliary information to the feature space, the proposed music recommendation model (NeuTrans) maps different information features to low-dimensional vector representation by non-linear neural networks. Specifically, the NeuTrans separately employs matrix factorization and attribute network embedding to extract auxiliary information (historical interaction, network structure and attributes of songs). The feature space of different information is obtained by nonlinearly mapping the feature space of the songs. Experimental analysis on two real-world datasets shows that our framework outperforms the state-of-the- art approaches on Top-N music recommendation.

Traffic flow prediction is a crucial task for the intelligent traffic management and control. Various machine learning based methods have been applied in this field. Most of these methods encounter three fundamental issues: feature representation of traffic patterns, learning from single location or network, and data quality. In order to address these issues, in this work we present a deep architecture for traffic flow prediction that learns deep hierarchical feature representation with spatio-temporal relations over the traffic network. Furthermore, we design an ensemble learning strategy via random subspace learning to make the model be able to tolerate incomplete data. The experimental results corroborate the effectiveness of the proposed approach compared with the state of the art methods.

With the growing popularity of online social platforms, it has been universally recognized that incorporating social relations into recommender systems can usually alleviate the problem of data sparsity. However, social recommender systems based on explicit relations are not as successful as expected due to the noise and the social cold issue of explicit social links. The intuition of utilizing explicit relations is that users share similar preferences if they are friends in the social network. In fact, quite a lot of users who are distant from each other in the social network also have similar tastes. The user item network and the user social network can provide useful information that can complement each other, so that exploring the implicit friends using the heterogeneous network they formed would be more helpful. In this paper, we propose an approach IFSR to discover implicit friends over the heterogeneous network to improve the performance of social recommendation. To find out reliable implicit ties, we first model the system as a heterogeneous network upon which both the preferences and social information are coupled. Over the HIN, similarities between each pair of users can be quantified through network embedding based representation learning. To reduce the computational cost while preserving the information embedded in the original networks and uncover the latent information hiding in the HIN, several meaningful meta-paths over the HIN are designed to guide the process of random walks. Finally, the Top-K implicit friends are incorporated into a social bayesian ranking model to enhance the performance of item ranking. Experimental results on three datasets demonstrate IFSR outperforms the state-of-the-art methods and illustrate why the implicit friends are advantageous for social recommendation.

Traffic flow prediction is a fundamental component in intelligent transportation systems. However, many existing prediction models endure several shortages. Most of the methods are constructed as a shallow model, which is difficult to reveal the intrinsic spatio-temporal relations embedded in traffic raw data. Moreover, the separation of feature learning and predictor learning brings a sacrifice of model performance. Then the hand designed features are difficult to be tuned appropriately. Finally, few existing methods consider the incomplete data problem which is in fact very severe for practical application. In this work, we develop a deep learning model to predict traffic flows. The main contribution is development of an architecture that integrates random subspace learning and ensemble learning on deep convolutional neural networks. The proposed model takes the traffic flow data as an image, and considers both exploring spatio-temporal correlations in the unified architecture and the incomplete data problem. The experimental results, using traffic data originated from the California Freeway Performance Measurement System (PeMS), corroborate the effectiveness of the proposed approach compared with the state of the art.

The new cyberattack pattern of advanced persistent threat (APT) has posed a serious threat to modern society. This paper addresses the APT defense problem, that is, the problem of how to effectively defend against an APT campaign. Based on a novel APT attack-defense model, the effectiveness of an APT defense strategy is quantified. Thereby, the APT defense problem is modeled as an optimal control problem, in which an optimal control stands for a most effective APT defense strategy. The existence of an optimal control is proved, and an optimality system is derived. Consequently, an optimal control can be figured out by solving the optimality system. Some examples of the optimal control are given. Finally, the influence of some factors on the effectiveness of an optimal control is examined through computer experiments. These findings help organizations to work out policies of defending against APTs.

The concept of smart grid gained tremendous attention among researchers and utility providers in recent years. How to establish a secure communication among smart meters, utility companies, and the service providers is a challenging issue. In this paper, we present a communication architecture for smart grids and propose a scheme to guarantee the security and privacy of data communications among smart meters, utility companies, and data repositories by employing decentralized attribute based encryption. The architecture is highly scalable, which employs an access control Linear Secret Sharing Scheme (LSSS) matrix to achieve a role‐based access control. The security analysis demonstrated that the scheme ensures security and privacy. The performance analysis shows that the scheme is efficient in terms of computational cost.

Prediction of Quality of Service (QoS) values plays an important role in service selection, discovery, and recommendation. Previous works show that the QoS values would be influenced by the location information. However, these researches do not consider the fact that some users may provide untrustworthy QoS values even though they are in the same location region. QoS values from these unreliable users could significantly affect the QoS prediction accuracy. To address this issue, this paper proposes an alternative and efficient approach to predict the missing QoS values, referred as the Location and Reputation aware Matrix Factorization based Location Information (LRMF). LRMF combines both the user's reputation and location information to achieve more accurate prediction results. Experiments are conducted on a real-world Web service QoS dataset, and results show that the proposed method outperforms many other existing QoS prediction methods.

Traditional recommender systems assume that all the users are independent, and they usually face the cold start and data sparse problems. To alleviate these problems, social recommender systems use social relations as an additional input to improve recommendation accuracy. Social recommendation follows the intuition that people with social relationships share some kinds of preference towards items. Current social recommendation methods commonly apply the Matrix Factorization (MF) model to incorporate social information into the recommendation process. As an alternative model to MF, we propose a novel social recommendation approach based on Euclidean Embedding (SREE) in this paper. The idea is to embed users and items in a unified Euclidean space, where users are close to both their desired items and social friends. Experimental results conducted on two real-world data sets illustrate that our proposed approach outperforms the state-of-the-art methods in terms of recommendation accuracy.

With the development of information technology, more and more web services have emerged, thereby making it difficult for customers to find their favorite services quickly and accurately. To overcome this difficulty, recently the collaborative filtering (CF) technique has been widely employed for personalized service recommendation, meanwhile improving the profits of service providers. Although the CF-based web service recommender systems have shown their potential, they appear to be vulnerable to shilling attack problems. Therefore, in this paper we analyze a general form of web service shilling attacks and four kinds of classical attack models, e.g., average attack, bandwagon attack, random attack, and segment attack are thoroughly investigated. Furthermore, we also study the impact of distribution-aware Pareto attack models. To demonstrate how shilling attacks alter the recommendation results, this paper analyzes 1) the variation of Quality-of-Service (QoS) prediction values of target services, 2) the QoS value prediction shifts of services with short response time which are more likely recommended, and 3) the comparison of prediction shift caused by classical attack models and Pareto attack models. The experimental results on WS-DREAM dataset revealed several interesting findings about the predictions of QoS values of target service correlated to different attack models. It is expected that this work can provide some insight for future vulnerability analysis of CF-based web service recommender systems.

While massive open online courses (MOOCs) have gained increasing popularity in recent years, dropout prediction has been an important task to solve due to the high rates of dropout students found in MOOCs. Current methods normally apply supervised learning methods to dropout prediction, using general features extracted from behavior records without the concern of behavior types. However, student learning behavior is diverse and there are no sufficient labeled data to train a model because it is a time-costing task to label enormous data in practice. To solve these problems, in this work, we proposed a novel multi-view semi-supervised learning model based on behavior features for the dropout prediction task. Specifically, we derived features from each type of learning behavior to form multi-view behavior features. In addition, based on these features, we proposed a new multi-view semi-supervised learning method to make use of a large number of unlabeled data to assist insufficient labeled data for improving prediction performance. We conducted experiments on KDD Cup 2015 dataset, and the results show that our proposed method achieves better prediction of student dropouts as compared to state-of-the-art approaches.

A functions distributed network, called universal learning networks with branch control of relative strength (ULNs with BR), is proposed. The point of the paper is to adjust the outputs of the intermediate nodes of the basic network using an additional branch control network. The adjustment multiplies the nodes outputs by the coefficients ranging from zero to one, which is obtained from the branch control network. Therefore, the following are characterized in ULNs with BR, (1) the branch is cut when the coefficient of its branch is zero, and (2) multiplication is carried out in the nodes outputs adjustment when the coefficient takes a nonzero value. ULNs with BR is applied to two-spirals problem. The simulation results show that ULNs with BR exhibits better performance than the conventional neural networks with comparable complexity.

Universal learning networks with branch control (BrcULNs) are proposed, which consist of basic networks and branch control networks. The branch control network can be used to determine which branches of the basic network should be connected or disconnected. This determination depends on the inputs or the network flows of the basic network. Therefore, by using the BrcULNs, locally functions distributed networks can be realized depending on the values of the inputs of the network or the information of the network flows. The proposed network is applied to some function approximation problems. The simulation results show that the BrcULNs exhibit better performance than the conventional networks with comparable complexity.