Collaborative Computing: Networking, Applications and Worksharing

When a natural disaster occurs, ground base stations (BSs) are destroyed and cannot provide communication services. Rapid restoration of communication is of great significance to the lives of trapped persons. This paper studies the problem of unmanned aerial vehicle (UAV) equipped with mobile edge computing (MEC) servers to provide communication and computing services for ground users in the scenario where the ground infrastructure is destroyed. We designed a UAV-to-Cell offloading system, which provides services in units of cells. By determining the hover locations (HLs) and trajectories, the UAV can handle more tasks with limited battery energy. Since tasks have time limit requirements, the order of processing will affect the task data size of the system. We solve this problem by joint cell clustering and path planning. Among them, elliptic clustering is used to divide the cells, the 3D position of the UAV is determined according to the quality of user service, and the double deep Q-network (DDQN) algorithm is used to determine the trajectory of the UAV. Simulation experiments demonstrate the effectiveness and efficiency of our proposed strategy by comparing it with the baselines.

To better understand complex human emotions, there is growing interest in utilizing heterogeneous sensory data to detect multiple co-occurring emotions. However, existing studies have focused on extracting static information from each modality, while overlooking various interactions within and between modalities. Additionally, the label-to-modality and label-to-label dependencies still lack exploration. In this paper, we propose LAbel-induced Mixed-level Blending (LAMB) to address these challenges. Mixed-level blending leverages shallow but manifold self-attention and cross-attention encoders in parallel to model unimodal context dependency and cross-modal interaction simultaneously. This is in contrast to previous works either use one of them or cascade them successively, which ignores the diversity of interaction in multimodal data. LAMB also employs label-induced aggregation to allow different labels to attend to the most relevant blended tokens adaptively using a transformer-based decoder, which facilitates the exploration of label-to-modality dependency. Unlike common low-order strategies in multi-label learning, correlations among multiple labels can be learned by self-attention in label embedding space before being treated as queries. Comprehensive experiments demonstrate the effectiveness of our methods for multimodal multi-label emotion detection.

In Visual Question Answering (VQA) task, extracting semantic information from multimodalities and effectively utilizing this information for interaction is crucial. Existing VQA methods mostly focus on attention mechanism to reason about answers, but do not fully utilize the semantic information of modalities. Furthermore, the question and the image relation description through attention mechanism may cover some conflicting information, which weakens multi-modal semantic information relevance. Based on the above issues, this paper proposes a Multi-layer Semantics Awareness Model (MSAM) to fill the lack of multi-modal semantic understanding. We design a Bi-affine space projection method to construct multi-modal semantic space to effectively understand modal features at the semantic level. Then, we propose to utilize contrastive learning to achieve semantic alignment, which effectively brings modalities with the same semantics closer together and improves multi-modal information relevance. We conduct extensive experiments on the VQA2.0 dataset, and our model boosts the metrics even further compared to the baseline, improving the performance of the VQA task.

In the cyber attack and defense process, the opponent’s strategy is often dynamic, random, and uncertain. Especially in an advanced persistent threat scenario, it is not easy to capture its behavior strategy when confronted with a long-term latent, highly dynamic and unpredictable opponent. FlipIt game can model the stealth interaction of advanced persistent threat. However, it is insufficient for traditional reinforcement learning approach to solve real-time and non-stationary game model. Therefore, how to model a non-stationary opponent implicitly and keep the defense agent’s advantage continuously is essential. In this paper, we propose an extended FlipIt game model incorporating opponent modeling. And then we propose an approach that combines deep reinforcement learning, opponent modeling, and dropout technology to perceive the behavior of a non-stationary opponent and defeat it. Instead of explicitly identifying the opponent’s intention, the defense agent observes the opponent’s last move actions from the game environment, stores the information in its knowledge, then perceives the opponent’s strategy and finally makes a decision to maximize its benefits. We show the excellent performance of our approach whether the opponent adopts traditional, random or composite strategies. The experimental results demonstrated that our approach can perceive the opponent quickly and maintain the superiority of suppressing the opponent.

Mobile Edge Computing (MEC) is a technology that enables on-demand the provision of computing and storage services as close to the user as possible. In an MEC environment, frequently visited content can be deployed and cached upon edge servers to boost the efficiency of content delivery and thus improving user-perceived experience. However, due to the dynamic nature of MEC, it remains a great challenge how to fully exploit mobility information in yielding high-quality content caching decisions for delay-sensitive real-time mobile applications. To address this challenge, this paper proposes a novel mobility-aware caching method by leveraging a Multi-Agent Deep Reinforcement Learning-Based (MAACC) Approach model. The proposed method synthesizes a content fitness algorithm for estimating the priority of caching content with high user fitness and a collaborative caching strategy built upon a multi-agent deep reinforcement learning model. Empirical results clearly show that MAACC outperforms its peers regarding cache hit rate and transfer delay time.

Imbalanced datasets often result in poor predictive model performance. To address this, minority class sample expansion is used, but two challenges remain. The first is to use algorithms to learn the main features of minority class samples, and the second is to differentiate the generated data from the majority class samples. To tackle these challenges in binary classification, we propose the Discriminant-Autoencoder (D-AE) algorithm. It has two mechanisms based on our insights. Firstly, an autoencoder is used to learn the main features of minority class samples by reconstructing the data with added noise. Secondly, a discriminator is trained on the raw data to distinguish the generated data from the majority class samples. Our proposed loss function, Discriminant-\(L_\theta \), balances the discriminant and reconstruction losses. Results from experiments on three datasets show that D-AE outperforms baseline algorithms and improves dataset applicability.

Most methods of emotional dialogue generation focus on how to make the generated replies express the set emotion categories, while ignoring the control over the semantic content of the replies. To this end, in this paper, we propose a emotion- and content-controllable response generation model, ECCRG. ECCRG allows for text-controlled conditions and integration into the decoding process of the language model through a self-attention layer, enabling more precise control over the content of the generated responses. We use a variety of optimization objectives including self-reconfiguration loss and adversarial learning loss to jointly train the model. Experimental results show that ECCRG can embody the set target content in the generated responses, allowing us to achieve controllability on both emotion and textual content.

Origin-Destination (OD) Matrix Prediction is an important part of public transportation service which aims to predict the number of passenger demands from one region to another and capture the passengers’ mobility patterns. This problem is challenging because it requires forecasting not only the number of demands within a region, but the origin and destination of each trip as well. To address this challenge, we propose an effective model, ODCRN (Origin-Destination Convolution Recurrent Network) which incorporates traffic context and bi-directional semantic information. First, we obtain the semantic embedded features of the region as the static traffic context by the Node2vec algorithm, and the traffic flow of the region is counted as the dynamic traffic context. Second, we construct two adjacency matrices which represent origin-destination and destination-origin travel demands within urban areas respectively based on the OD matrices of each time slot, and use the graph convolutional network to aggregate traffic context information of the semantic neighbors in both directions. Then, we use a unit constructed by GRU and the graph convolution network to capture the spatial-temporal correlations of the input data. Finally, we use those correlations and traffic contexts to predict the OD matrix for the next time slot. Our model is evaluated on TaxiNYC and TaxiCD datasets, and experimental results demonstrate the superiority of our ODCRN model against the state-of-the-art approaches.

The accurate segmentation of stroke lesion regions holds immense significance in shaping treatment strategies and rehabilitation protocols. Due to the large difference in the volume of stroke lesion areas and the great similarity between lesion areas and normal tissues, most of the existing methods for lesion segmentation cannot deal with these problems well. This paper proposes a novel network named MD-TransUNet for the segmentation of stroke lesions, whose framework is based on the UNet architecture. To fully obtain deep image features, it uses ResNet50 for downsampling. MD (multi-dilated) module is employed as the skip connection to gain more receptive fields. Different receptive fields can adapt to varying volumes of lesion areas. Then, a feature extraction module with multi-level attention mechanism is designed using ConvLSTM, non-local spatial attention, and channel attention modules to suppress useless information expression in skip connections and upsampling processes while focusing more on effective spatial and channel information in features. The experiments show that our proposed network gets superior performance than benchmark methods and indicates the generalization and effectiveness of the proposed model.

Internet of Things (IoT) is network based on information carriers such as the Internet and traditional telecommunications networks, so that all ordinary physical objects that can be independently addressed can be interconnected. In the face of the IoT produces a large of time series data, which is very necessary to detect anomaly data. Transformer has proven to be a powerful tool in several areas, but still has some limitations, such as the prediction accuracy is not high enough. As the dominant trend of multivariate time series in different scenarios becomes increasingly evident, it is particularly important to accurately capture the spatio-temporal features between them. To address these issues, we propose Dynamic Graph transFormer (DGFormer), an effective Dynamic Graph Transformer based Anomaly Detection Model for IoT Time Series. We first use Transformer with anomaly attention mechanism to extract time features. Then, a dynamic relationship embedding strategy is proposed to capture spatio-temporal features dynamically and learn the adjacency matrix adaptively. Besides, each layer of GNN is soft clustered by Diffpooling. Finally, in order to further improve the detection performance of model, we integrate the traditional autoregressive linear model with the nonlinear neural network in parallel. The experimental results show that the proposed model achieves the highest F1-score on three public IoT datasets, and the F1-score is improved by 19.3% on average.

Gesture recognition plays a pivotal role in enabling natural and intuitive human-computer interaction (HCI), finding applications in diverse domains such as smart homes, robot control, and virtual reality. Thanks to advances in computer vision, the most popular method currently is to use the camera for gesture recognition. However, the camera struggles to function properly in poor lighting and inclement weather, and risks invading privacy. Due to the robust and non-invasive features of millimeter-wave radar, gesture recognition based on millimeter-wave radar has received extensive attention from researchers in recent years. In this paper, we propose a novel graph neural network named STAPointGNN for gesture recognition using millimeter-wave radar. In order to better extract features in the spatial and temporal dimensions of point clouds collected by millimeter-wave radar, we designed a spatial-temporal attention mechanism based on graph neural network. We also propose a novel point flow embedding method to capture the motion features of the point clouds in adjacent frames. To verify the superiority of our method, we conduct experiments on two public millimeter-wave radar gesture recognition datasets. The results show that our model outperforms existing mainstream algorithms.

The massive volume and the inherent imbalance of graphs are inevitable challenges for efficient graph computing, primarily due to the limited capacity of main memory (DRAM). Fortunately, a promising solution has emerged in the form of hybrid memory systems (HMS) which combine DRAM and persistent memory (PMEM) to enable data-centric graph computing. However, directly transitioning existing DRAM-based models to HMS can lead to inefficiency issues, especially when crossing Non-Uniform Memory Access (NUMA) nodes. In this paper, we present NPGraph, a novel approach that fully exploits the advantages of HMS for in-memory graph computing models. The main contributions of NPGraph lie in three aspects. Firstly, a dual-block graph representation strategy is devised to accelerate the process of subgraph construction. By utilizing data layering, it fully utilizes the storage architecture of HMS and optimizes the data access process. Secondly, an adaptive push-pull update strategy is proposed to optimize the message-updating process. With data-driven algorithms, it dynamically migrates subgraphs which are used in future iterations. Thirdly, the effectiveness of NPGraph is evaluated on five public graph data sets. Our model can improve the temporal locality and the spatial locality of graph computing concurrently. Extensive evaluation results show that NPGraph outperforms state-of-the-art graph computing models by 21.67%–32.03%.

Entity prediction and relation prediction are the two major tasks of temporal knowledge graph (TKG) reasoning. The key to answering queries about future events is to understand historical trends and extract the information most likely to affect the future, i.e., the TKG reasoning task is both influenced by the trends of time-evolving graphs and directly driven by the facts relevant to a specific query. Existing methods mostly build models separately for these two characteristics, namely evolution representation learning and query-specific methods, failing to integrate these two crucial factors that determine reasoning results into a single framework. In this paper, we propose a novel temporal hybrid reasoning network (tHR-NET), simultaneously considering the modeling of graph feature space evolution and the enhancement of query-related feature representations in TKG. Specifically, we introduce a global graph space evolution module to extract graph trends, which influence entity/relation representations at each timestamp through a temporal view projection. Additionally, we propose a query-specific increment module for targeted enhancement of entity and relation representations, capturing query-related factors over extended durations. Through extensive experiments on real datasets, tHR-NET demonstrates distinct advantages in parallel entity and relation prediction.

As the field of code intelligence continues to grow, Code representation learning has emerged as a research hot spot. Given that code structure can be naturally represented as graphs, Graph Neural Networks (GNNs) have proven highly effective for learning graph representations of source code. Pooling, as an essential operation for GNN-based models, is limited in its ability to leverage the rich hierarchical information presented in tree-like graph, especially Abstract Syntax Trees. In order to learn the graph representation of code more effectively, we propose a novel pooling method called TreePool. TreePool directly splits tree-like graphs using depth filtering based on the tree structure to form a sequence of pooled graphs sorted by descending size of subgraphs. Then local-local contrastive learning between these neighboring subgraphs is conducted to preserve the information of the graph before pooling. Through TreePool, multiple views of representation are learned and fused to obtain the final code graph representation. We conduct TreePool on a supervised framework and experimental results demonstrate that the average improvements on two real-world datasets in terms of accuracy are 1.1% and 3.3%. It also exhibits excellent performance in an unsupervised framework. Our results show that TreePool can effectively learn meaningful Abstract Syntax Tree representation of code and exhibit good performance in code classification tasks.

In IoT systems, traditional encryption can be used to protect IoT applications and data at rest or in transit that transforms data in to ciphertext making it unreadable. However, it is very challenging to protect IoT systems against attacks targeting data and applications in use. Using homomorphic encryption, this work proposed a lightweight collaborative computing scheme to protect both applications and data in IoT environment that includes IoT devices, mobile apps, and cloud server. A novel key management system scheme proposed as a trusted third party to collaboratively generate and distribute keys by cloud servers and IoT devices, in which data is only visible to the data owner but keep encrypted to other parties. A SEAL-CKKS scheme and a K-means clustering algorithms were validated, and the experimental results demonstrated the effectiveness of proposed schemes, in which the K-means clustering algorithm in the plaintext state, the proposed scheme still maintains an accuracy up to 84.1%.

Writing high-quality unit tests plays a crucial role in discovering and diagnosing early-stage errors and preventing their further propagation throughout the development cycle. However, the low readability of existing automated test case tools hinders developers from directly using them. In addition, current approaches exhibit sensitivity to individual words in the input code, often producing completely different results for minor changes in the input code. To tackle these problems, we propose AssertGen, a powerful Java assertion generation model that maintains consistent output for minor variations in code snippets. Inspired by software mutation testing, we propose 11 heuristic strategies for code mutation, aiming to generate variant code that is human-readable but misleading to the model, by making minor changes to code text or structural information. Then, we use the variant code to attack the model to test the model’s robustness. We observe that the variant based on variable names (VM), the mutation based on method names (FM), and the mutation method False_Control_Flow, which adds additional control flow, have the greatest impact on the quality of generated assertions by the model. To enhance the robustness of AssertGen, we use multiple mutations to expand the original dataset, allowing the model to learn how to counter the instability caused by mutations during the training process. Experiment results show our assertion generation model achieves a BLEU score of 60.08 and a perfect prediction rate of 47.91%, surpassing previous work significantly.

Aiming at the issues of low comprehensiveness and timeliness of data, difficulty in balancing data anonymity and traceability, and challenges of securely storing massive data in traditional traffic data sharing systems, this paper proposes a UAV-VANET integrated system (UVIS) based on consortium blockchain. The UAV integrated into the VANET can promptly provide drivers and traffic managers with comprehensive traffic information and images for traffic planning, thus enhancing transportation efficiency and safety. To achieve traceability of anonymous data sharing, we introduce a proxy re-encryption mechanism to realize precise data access control, which can not only protect data and identity privacy but also trace the true identity of malicious users. Additionally, it effectively prevents the collusion between proxies and data requesters from stealing unauthorized confidential information. To alleviate the pressure of traffic data storage, we adopt a storage method that combines blockchain and IPFS, ensuring secure storage of massive data. Security analysis shows that the UVIS has achieved secure sharing of traffic data. We analyze its efficiency theoretically, and demonstrate the practicality through experiments.

With the development of information and communication technologies, the services provided by smart grid attract more users to join smart grid. However, with the explosive growth of the number of smart meters, the transmission between the control center and smart meters has brought huge data transmission and computing costs to the smart grid, which is prone to network congestion, untimely power service supply and other network conditions. This paper proposes a Physically Unclonable Function (PUF)-based lightweight group authenticated metering data collection scheme with privacy protection in smart grid (PGAC). The PGAC scheme is designed with lightweight cryptographic primitives, which is suitable for resource-constrained devices. In addition, the PGAC scheme divides the users into groups and uses the gateway as the repeater and aggregator of the communication data of each group, which reduces signaling and communication costs for activating additional request messages from a large number of devices. Security analysis shows that the PGAC scheme maintains the security and privacy of the data collection process for large-scale smart meters. Functional analysis, theoretical analysis and performance analysis show that PGAC scheme has better authentication function and low communication cost.

The rapid growth of malware and its variants has a significant detrimental effect on the security of the Internet infrastructure. In recent years, deep learning-based methods have demonstrated significant success in malware detection. Nonetheless, there are concerns regarding the requirement for substantial labeled data and the feature selection methods used in present approaches. In this paper, we propose a semi-supervised learning-based method for malware traffic classification, which exploits the raw bitmap representation of malware traffic. We employ stacked bi-LSTM to learn the feature representation of malware traffic and adopt semi-supervised learning (SSL) to enhance the model performance by leveraging unlabeled traffic. Pseudo-labeling and consistency regularization are used to produce pseudo-labels, which can compute unsupervised loss. The loss function consists of two terms: a supervised loss applied to labeled data and an unsupervised loss, which are combined together for model training. Experiments indicate that our method is capable of classifying malware traffic with satisfactory accuracy.

In large-scale distributed computing systems, coded computing has attracted considerable attention since it can effectively mitigate the impact of stragglers. Nonetheless, several emerging issues seriously restrict the performance of coded distributed systems. First, the presence of colluding workers collude results in serious privacy leakage issues. Second, few existing works consider security issues in data transmission. Third, the number of required results to wait for increases with the degree of polynomial functions. In this paper, we propose a secure and private approximated coded distributed computing (SPACDC) scheme that addresses the aforementioned issues simultaneously. The SPACDC scheme ensures data security during the transmission process by leveraging a proposed matrix encryption algorithm based on elliptic curve cryptography. Unlike existing coding schemes, our SPACDC scheme does not impose strict constraints on the minimum number of results required to wait for. Furthermore, the SPACDC scheme provides information-theoretic privacy protection for raw data. Finally, extensive performance analysis is provided to demonstrate the effectiveness of the proposed SPACDC scheme.

Internet of Things (IoT) is an evolving paradigm for building smart cross-industry. The data gathered from IoT devices may have anomalies or other errors for various reasons, such as malicious activities or sensor failures. Anomaly detection is thus in high need for guaranteeing trustworthy execution of IoT applications. Existing IoT anomaly detection methods are usually built upon unsupervised methods and thus can be inadequate when facing complex IoT data regularity. In this article, we propose a semi-supervised approach for detecting IoT time series anomalies based on Graph Structure Learning (GSL) using multi-layer perceptron Graph Convolutional Networks (GCN) and the Mean Teachers (MT) mechanism. The proposed model is capable of leveraging a small amount of labeled data (1% to 10%) to achieve high detection accuracy. We adopt Mean Teachers to utilize unlabeled data for enhancing the model’s detection performance. Moreover, we design a novel graph structure learning layer to adaptively capture the IoT data features among different nodes. Experimental results clearly suggest that the proposed model outperforms its competitors on two public IoT datasets, achieving 82.85% in terms of F1 score and 22.8% increase.

As code intelligence and collaborative computing advances, code representation models (CRMs) have demonstrated exceptional performance in tasks such as code prediction and collaborative code development by leveraging distributed computing resources and shared datasets. Nonetheless, CRMs are often considered unreliable due to their vulnerability to adversarial attacks, failing to make correct predictions when faced with inputs containing perturbations. Several adversarial attack methods have been proposed to evaluate the robustness of CRMs and ensure their reliable in application. However, these methods rely primarily on code’s textual features, without fully exploiting its crucial structural features. To address this limitation, we propose STRUCK, a novel adversarial attack method that thoroughly exploits code’s structural features. The key idea of STRUCK lies in integrating multiple global and local perturbation methods and effectively selecting them by leveraging the structural features of the input code during the generation of adversarial examples for CRMs. We conduct comprehensive evaluations of seven basic or advanced CRMs using two prevalent code classification tasks, demonstrating STRUCK’s effectiveness, efficiency, and imperceptibility. Finally, we show that STRUCK enables a more precise assessment of CRMs’ robustness and increases their resistance to structural attacks through adversarial training.

The data of applications in connected and autonomous vehicles are important, which is usually collected by service providers to improve their services, such as object detection model. But, wireless communication is susceptible to various kinds of attacks. Thus, the data of the application module needs to be securely shared to the corresponding service provider. However, current schemes are with limited performance while a service provider collects multiple application data at the same time. By adopting signcryption and chaotic map, an efficient authentication and key agreement scheme is proposed, while batch authentication is achieved for efficient message authentication of multiple applications, and the efficient revocation is realized based on Chinese remainder theorem under the assistance of trusted execution environment supported vehicle computing/communication unit. The formal security proof shows that the scheme is secure under the random oracle model, and the experiment results shows that the scheme is more efficient than related schemes and can meet the requirements of CAV.

Measuring the similarity between two business process models holds significant importance across various applications. At present, there are many different similarity calculation methods, such as structural similarity based on the graph edit distance(GED), text similarity based on task node description, and behavioral similarity calculation based on path matching. However, existing similarity computation methods cannot produce reliable results since: (1) To apply GED, business process graphs will be simplified to homogeneous graph where the heterogeneity as well as the routing semantics of the business process is removed. (2) To derive comprehensive similarity evaluation, linear weighted sum of different similarity metrics is a common way, but the final result strongly depends on the weighting coefficients that are empirically assigned. In this paper, we fuse multidimensional metrics to compensate for the sole reliance on structural similarity based on GED. To address the limitations of comprehensive evaluation, we propose a novel multidimensional process similarity evaluation method based on the entropy weight method and the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) method. We also design a experimental method to verify the effectiveness of our method, leveraging an open source dataset. The experiment shows that our method can better represent the similarity of business process graphs than other methods.

Efficient image retrieval and recognition are pivotal for optimal mobile web vision services. Traditional web-based solutions offer limited accuracy, high overhead, and struggle with vast image volumes. Transferring images for real-time cloud recognition demands stable communication, and large-scale concurrent requests strain computational and network resources. This paper introduces a distributed recognition approach, leveraging cloud-edge-device collaboration through edge computing’s low latency and high bandwidth. We present a lightweight image saliency detection model tailored for mobile web, enhancing initial image feature extraction. Additionally, we introduce an edge-based, deep learning-driven method to amplify image retrieval speed and precision. We incorporate a location and popularity-based caching system to alleviate strains on cloud resources and network bandwidth during extensive image requests. Our real-world tests validate our approach: our saliency detection model outpaces the benchmark by reducing the model size by up to 94%, making it suitable for mobile web deployment. The proposed method improves retrieval accuracy by 40% over cloud-based counterparts and cuts response latency by over 60%.

In finger-vein recognition tasks, obtaining large labeled datasets for supervised deep learning is often difficult. To address this challenge, self-supervised learning (SSL) provides a solution by first pre-training a neural network using unlabeled data and subsequently fine-tuning it for downstream tasks. Contrastive learning, a variant of SSL, enables effective learning of image-level representations. To address the issue of insufficient labeled data for vein feature extraction and classification, we propose CL3A-FV, a Contrastive Learning-based Finger-Vein image recognition approach with Automatic Adversarial Augmentation in this paper. Specifically, CL3A-FV consists of the dual-branch augmentation network, Siamese encoder, discriminator, and distributor. The training process involves two steps: 1) training the Siamese encoder by updating its parameters while keeping other components fixed; and 2) training the dual-branch augmentation network with a fixed Siamese encoder, integrating a discriminator to distinguish views generated by the two branches, and a distributor to constrain the distribution of the augmented data. Both networks are updated adversarially using the stochastic gradient descent. We conduct extensive experiments to evaluate CL3A-FV on three finger-vein datasets, and the experimental results show that the proposed CL3A-FV achieves significant improvements compared to traditional self-supervised learning techniques and supervised methods.

Quality of service (QoS) is the main factor in service selection and recommendation, and it is influenced by dynamic factors, such as network condition and user location, and static factors represented by the invocation sequence at a fixed time slice. In order to jointly consider these two factors, this work proposes a multi-dimensional sequential contrastive learning framework named MDSCL, which applies contrastive learning method to learn the sequence representations of both user and time dimensionalities. An overlap crop augmentation strategy is proposed to obtain positive examples for user sequences and time sequences, respectively. Besides, MDSCL includes an integrated feature extractor that combines WaveNet and BiLSTM to facilitate the long short-term feature capturing. Extensive experiments on WSDREAM have been conducted to verify the effectiveness of our approach.