Computer Engineering and Applications

Select

Active Microservice Fine-Grained Scaling Algorithm

PENG Kai, MA Fangling, XU Bo, GUO Jialu, HU Menglan

Computer Engineering and Applications 2024, 60 (8): 274-286. DOI: 10.3778/j.issn.1002-8331.2303-0543

Abstract （13）

PDF（pc）（876KB）（23）

Save

Microservice architecture has become the basic service architecture of cloud data center. However, the existing studies on the elastic scaling scheme of microservice systems are mostly based on horizontal scaling at the service or instance level, ignoring the fine-grained vertical scaling that can make full use of single server resources, resulting in resource waste. Therefore, an active microservice fine-grained elastic scaling algorithm is designed in this paper. The algorithm forecasts the request arrival rate to preconfigure the system resources. Based on the predicted results, the square root staffing rule is applied to calculate the number of required resources, and then the microservice is scaled by using the fine-grained resource control feature of vertical scaling and the high availability of horizontal scaling. Finally, an instance migration algorithm based on microservice dependency is applied to further reduce the resource overhead. Experimental results show that the proposed algorithm is effective in optimizing the delay and overhead of microservice systems.

Reference | Related Articles | Metrics

Select

Multi-View Group Recommendation Integrating Self-Attention and Graph Convolution

WANG Yonggui, WANG Xinru

Computer Engineering and Applications 2024, 60 (8): 287-295. DOI: 10.3778/j.issn.1002-8331.2304-0035

Abstract （18）

PDF（pc）（604KB）（25）

Save

In order to solve the problem that most existing group recommendations only learn group representation from a single interaction between the group and the user, and that the fixed fusion strategy is difficult to dynamically adjust the weight. A multi-view group recommendation model (MVGR) is proposed, which integrates self-attention and graph convolution. Three different views, member level, item level and group level, are designed to capture high-level collaborative information among groups, users and items, alleviate the problem of data sparsity, and enhance group representation modeling. For item level views, the graph convolution neural network based on dichotomous graph is used to learn group preference vector and item embedding. MVGR further proposes an adaptive fusion component to dynamically adjust different view weights to get the final group preference vector. Experimental results on two real dataset show that the hit ratio (HR) and normalized discounted cumulative gain (NDCG) of the MVGR model are improved by an average of 8.89?percentage points and 1.56 percentage points on the Mafengwo dataset, and by an average of 2.79 percentage points and 2.7 percentage points on the CAMRa2011 dataset compared to the baseline model.

Reference | Related Articles | Metrics

Select

Cooperative Task Processing in Multi-UAV Assisted Mobile Edge Computing

CAO Huijuan, YU Genghua, CHEN Zhigang

Computer Engineering and Applications 2024, 60 (4): 298-305. DOI: 10.3778/j.issn.1002-8331.2301-0145

Abstract （56）

PDF（pc）（2281KB）（42）

Save

Mobile edge computing (MEC) is considered as a key technology to deal with computing intensive and delay sensitive tasks. However, in disaster response, emergency rescue and other scenarios, it is impossible to quickly deploy edge servers and provide services. Unmanned aerial vehicle (UAV) assisted MEC has attracted much attention because of its simple deployment and strong mobility. However, the computing resources and energy of UAV are limited, how to allocate resources becomes a difficult problem. To solve this problem, a strategy of efficient utilization resources with multi-UAV cooperation (LUAVs-Cor) is proposed. This strategy dynamically processes tasks through the multi-UAV cooperation. In order to make full use of computing resources of UAVs, the task transmission strategy is determined by searching for the optimal task combination. In addition, the number of UAVs dispatched is optimized according to the UAV processing capacity, the number of tasks and the processing status of tasks, which achieves dynamic deployment of UAVs and reduces energy consumption. By simulation experiments, it is concluded that the service capacity of LUAVs-Cor strategy has been increased by about 6.8%, and the overall energy consumption of UAV has been reduced by 10.3%. LUAVs-Cor strategy has a low collaboration cost and serves more users at a lower energy cost.

Reference | Related Articles | Metrics

Select

Multi-Task Joint Learning for Graph Convolutional Neural Network Recommendations

WANG Yonggui, ZOU Heyu

Computer Engineering and Applications 2024, 60 (4): 306-314. DOI: 10.3778/j.issn.1002-8331.2303-0508

Abstract （53）

PDF（pc）（2409KB）（35）

Save

Collaborative filtering recommendation based on graph neural network can mine the interaction information between users and items more effectively, but its performance is still affected by the problems of sparse data and low quality of representation learning. Therefore, a multi-task joint learning model for graph convolutional neural network recommendation (MTJL-GCN) is proposed. Firstly, the graph neural network is used to gather the homogeneous structural information on the user-item interaction graph and form the structural neighbor relationship with the initial embedding information, and the comparative learning assistance task of node neighbor relationship is designed to alleviate the data sparse problem. Secondly, random unified noise is added to the original representation of nodes to enhance the representation-level data, and a comparative learning assistance task is constructed for the node representation relationship. The learning objectives of alignment and uniformity are proposed to improve the quality of representation learning. Finally, the graph collaborative filtering recommendation task is combined with the comparative learning assistance task and the direct optimization of learning objectives for joint training, so as to improve the recommendation performance. Experiments on two public datasets of Amazon-books and Yelp2018 show that the model performs better than the baseline model in the two recommended performance indexes of Recall@k and NDCG@k, which proves the validity of the MTJL-GCN model.

Reference | Related Articles | Metrics

Select

Recommendation Model Based on Time Aware and Interest Preference

TANG Pan, WANG Xueming

Computer Engineering and Applications 2023, 59 (24): 268-276. DOI: 10.3778/j.issn.1002-8331.2212-0386

Abstract （47）

PDF（pc）（644KB）（32）

Save

To address the problem that traditional recommendation models cannot mine users’ fine-grained interest preferences, a recommendation model based on time aware and interest preferences（TAIP） is proposed. In TAIP model, the time interval information of user interaction is introduced into the sequence embedding matrix as auxiliary information, and a multi-scale temporal convolutional network with channel and spatial attention mechanisms is designed to accurately extract fine-grained short-term preferences. At the same time, the Transformer encoder is used to mine long-term preferences between target items and user interests. Finally, a fully connected network is used to achieve global feature fusion to provide recommendations. Experiments are conducted on the publicly available datasets MovieLens-1M and YELP. Compared with other models, TAIP model improves at least 4.84% and 1.38% in HR, NDCG and MRR, which has better recommendation performance and verifies the effectiveness of TAIP model.

Reference | Related Articles | Metrics

Select

CodeBERT Based Code Classification Method

CHENG Siqiang, LIU Jianxun, PENG Zhenlian, CAO Ben

Computer Engineering and Applications 2023, 59 (24): 277-288. DOI: 10.3778/j.issn.1002-8331.2209-0402

Abstract （62）

PDF（pc）（702KB）（52）

Save

With the continuous development of code big data, the amount of source code in the code base is gradually growing, which makes software code management more complex. How to quickly and effectively classify and manage the code in the code base is of great importance to the development of software engineering. The article introduces pre-trained models to code classification research for the first time and proposes an optimized code classification method, CBBCC, which firstly uses wordpiece to pre-process the source code. Secondly, a CodeBERT pre-training model is used to characterise the source code. Finally, the classification task is fine-tuned on the basis of the pre-trained model. To verify the effectiveness of the proposed model, experimental analysis is conducted on the POJ104 dataset. The experimental results show that the CBBCC model achieves more than 98% in all classification metrics compared to the seven benchmark models. The accuracy is improved by 1.1 percentage points over the current optimal model, reaching the SOTA value for the classification task on the POJ104 code classification dataset. CBBCC can effectively annotate code, improve the management of open source community source code and promote the development of the software engineering field.

Reference | Related Articles | Metrics

Select

Domain Adversarial Adaptive Short-Term Workload Forecasting Model

LIU Chunhong, JIAO Jie, WANG Jingxiong, LI Weili, ZHANG Junna

Computer Engineering and Applications 2023, 59 (24): 289-297. DOI: 10.3778/j.issn.1002-8331.2211-0036

Abstract （20）

PDF（pc）（817KB）（24）

Save

The accuracy of workload prediction is one of the main factors affecting the elastic resource management of cloud platforms. And there are a large number of short task workload sequences with insufficient historical information and unsmooth characteristics in the cloud, which makes it difficult to select appropriate models for accurate prediction. In this paper, a domain adversarial workload prediction model is proposed. The model uses SSA（singular spectrum analysis） to smooth the workload and solve the problem of irregularity. Similarity calculations are performed by combining MASS_V4（the fourth version of Mueen’s algorithm for similarity search） with temporal features to obtain suitable source-domain data-assisted migration prediction. The GRU（gated recurrent unit） is used as the reference to construct the network, a new loss function defined is with Y-discrepancy, and a prediction model is constructed with strong short-workload feature representation ability in the adversarial process. The proposed method is compared with other commonly used cloud workload prediction algorithms on two real cloud platform datasets and both show higher prediction accuracy.

Reference | Related Articles | Metrics

Select

Multi-Server Collaborative Task Caching Strategy in Edge Computing

MA Shixiong, GE Haibo, SONG Xing

Computer Engineering and Applications 2023, 59 (20): 245-253. DOI: 10.3778/j.issn.1002-8331.2211-0238

Abstract （67）

PDF（pc）（690KB）（58）

Save

Aiming at the contradiction between the limited computing and storage resources of the edge server and a large number of user task requests, this paper designs an edge computing task cache network architecture based on multi-server collaboration. In this architecture, the edge server can cache and execute user tasks in memory. The cached tasks are executed on the cloud. Combining the characteristics of time-varying user task requests and users in adjacent areas tending to request similar tasks, a multi-server task caching algorithm（MSAC） based on improved Soft Actor-Critic is proposed. The algorithm aims at minimizing the user’s average task execution delay. In order to avoid repeatedly choosing the same action and converge to a local optimum, a maximum entropy model is introduced to encourage the edge server to explore the optimal action. By designing an experience sharing mechanism, the task caching strategy is optimized by collecting and learning the experience of local edge servers and adjacent servers. Simulation results show that compared with the highest popularity algorithm, independent SAC algorithm, DQN algorithm, and genetic algorithm, the proposed MSAC algorithm has the best effect in reducing the average execution delay of user tasks.

Reference | Related Articles | Metrics

Select

Data Stream Classification Method Combining Micro-Clustering and Active Learning

YIN Chunyong, CHEN Shuangshuang

Computer Engineering and Applications 2023, 59 (20): 254-265. DOI: 10.3778/j.issn.1002-8331.2210-0230

Abstract （30）

PDF（pc）（1058KB）（31）

Save

Data stream classification is an important research component in data mining, but the problems of concept drift and expensive labeling in data streams pose a great challenge to classification. Most of the existing research work adopts online classification technology based on active learning, which alleviates the problems of concept drift and limited labels to a certain extent. However, these methods are less efficient for classification and ignore the problem of memory overhead. Aiming at these problems, a data stream classification method combining micro-clustering and active learning is proposed（CALC）. Firstly, a new active learning hybrid query strategy is proposed to measure the importance of each microcluster during maintenance by combining it with error-based representative learning. Secondly, a set of microclusters is dynamically maintained to accommodate the concept drift generated in the data stream. In addition, an inert microcluster-based learning approach is used to achieve classification of the data stream and to accomplish online updates of the cached microclusters. Finally, comparative experiments are conducted using three real datasets and three simulated synthetic datasets, and the results show that CALC outperforms existing data stream classification algorithms in terms of classification accuracy and memory overhead. Compared with the benchmark model ORSL, the classification accuracy of CALC has been improved to a certain extent, and the average accuracy of the six data sets has been increased by 5.07, 2.41, 1.04, 1.03, 3.47 and 0.64 percentage points, respectively.

Reference | Related Articles | Metrics

Select

Research of Metadata Management Method of Hierarchical Storage System Based on HDFS

LIU Xiaoyu, XIA Libin, JIANG Xiaowei, SUN Gongxing

Computer Engineering and Applications 2023, 59 (17): 257-265. DOI: 10.3778/j.issn.1002-8331.2211-0230

Abstract （66）

PDF（pc）（5727KB）（32）

Save

With the continuous expansion of the scale of high-energy physics（HEP） experiments and the increase of experimental complexity, researchers are facing the challenge of big data storage. Considering the cost, energy consumption, storage cycle and maintenance management, the tape libraries with?large storage capacity and low cost have become an indispensable choice for mass storage systems in the field of HEP. However, HDFS heterogeneous storage doesn’t support tape library storage, it cannot meet the high cost performance requirements of storage system for the persistence and backup process of massive experimental data in Hadoop platform of HEP. In view of the above problems, in order to build an HDFS hierarchical storage system that supports disk-tape storage, the tape layer files can be seamlessly integrated in HDFS, and provide users with a unified file system namespace. In this research, it first overviews the existing methods of distributed file system metadata management, and further designs an improved one to realize the unified metadata management for HDFS hierarchical storage system. This method redesigns the file metadata structure in memory to build a unified memory directory tree, and implements access management and reliability assurance to achieve centralized and unified management of tape file metadata.?Test results show that the metadata server implements unified management of file metadata on heterogeneous resources and provides efficient metadata operation. The storage tiering system based on this method has high reliability. When reading and writing files of different sizes, the read and write throughput is better than that of the traditional storage tiering system EOSCTA in the HEP field.

Reference | Related Articles | Metrics

Select

Incorporating Relational Awareness and Temporal Attention for Temporal Knowledge Graph Completion

XU Zhihong, MAO Chen, WANG Liqin, DONG Yongfeng

Computer Engineering and Applications 2023, 59 (17): 266-274. DOI: 10.3778/j.issn.1002-8331.2210-0266

Abstract （68）

PDF（pc）（3433KB）（51）

Save

To address the problem that most of the existing temporal knowledge graph completion methods embed time information into triples and rely on static knowledge graph completion means to learn entity features, which cannot holistically consider structural information and temporal information in the graph. This paper proposes a temporal knowledge graph completion model incorporating relational awareness and temporal attention（RATA）. On the one hand, by introducing graph convolutional network with relation-aware aggregation mechanism to integrate entity and relation features, relation-specific parameters can enhance the expression ability of the message function, and encapsulate richer neighborhood context information. On the other hand, it employs the long short-term memory network incorporating self-attention mechanism to learn global and local features in the sequence. The experimental results on ICEWS18, ICEWS14, YAGO and WIKI show RATA generally outperforms the baseline model on MRR, Hits@1, Hits@3, and Hits@10, and has better advantages on large-scale temporal datasets.

Reference | Related Articles | Metrics

Select

Design and Implementation of Tape Library Storage System Based on Microservice Architecture

LIU Xiaoyu, XIA Libin, JIANG Xiaowei, SUN Gongxing

Computer Engineering and Applications 2023, 59 (15): 253-263. DOI: 10.3778/j.issn.1002-8331.2206-0067

Abstract （79）

PDF（pc）（777KB）（56）

Save

The establishment of HDFS tiered storage system with tape storage layer is an important part of improving Hadoop ecosystem in the field of high energy physics（HEP）. However, the traditional tape storage management system in HEP（such as Castor and CTA） doesn’t support HDFS. In addition, with the rapid growth of the amount of HEP data, the continuous development of current Internet technology and rapid changes in user needs, the traditional tape storage management system gradually presents the problem such as system expansion, load balancing, development and maintenance cost increases sharply. This paper designs and develops a tape storage management system based on microservice architecture. The system supports HDFS disk storage and distributes tape library resource management, file transfer, tape read/write and so on functions to each service instance in the form of microservices to disperse service pressure. Furthermore, aiming at the problem of low efficiency of traditional load algorithm, the system proposes and implements a load balancing algorithm based on server response factor. It sorts severs according to the server response factor calculated by user-defined parameters to ensure that user requests are scheduled to the server with the highest response factor for processing. ?Experimental results show that the tape library storage system in this paper can meet the requirements of HDFS file hierarchical storage. Compared with polling algorithm, the system performance of archiving based on server response factor is more than 6% better, and that of extraction is more than 64% better. ?Compared with random algorithm, archiving performance is more than 9% better, and extraction performance is more than 64% better. The application results of the system indicates that compared with the traditional system, the microservice architecture can decouple components and balance distributed loads in the system, and is more convenient in system development and maintenance.

Reference | Related Articles | Metrics

Select

Session-Based Recommendation Based on Multi-Graph Neural Network Incorporating Social Information

LEI Jingsheng, LI Ran, YANG Shengying, SHI Wenbin

Computer Engineering and Applications 2023, 59 (15): 264-273. DOI: 10.3778/j.issn.1002-8331.2209-0083

Abstract （51）

PDF（pc）（647KB）（35）

Save

In recommender systems, users’ interest in items is dynamic and influenced by various factors such as their own historical behavior and friends’ behavior. It has been a challenge for recommendation algorithms to jointly model users’ dynamic interests and social relationships. In this paper, the dynamic interests of users are explored by partitioning their behaviors into session sequences and modeling them as global graphs. After that, a social relationship graph based on users’ social relationships is constructed, and then the influence of users’ social relationships is captured through graph attention networks to dynamically determine the influence of each friend, and users’ dynamic interests are combined with friends’ social influence to obtain the final recommendation results. The algorithm is validated on Douban, Delicious and Yelp datasets. Compared with the most available baseline model, the algorithm improves more than 6 percentage points on the Douban dataset and more than 3 percentage points on the Delicious and Yelp datasets, which proves the effectiveness of the algorithm.

Reference | Related Articles | Metrics

Select

Collaborative Filtering Recommendation Algorithm Based on Graph Convolution Attention Neural Network

WANG Wei, DU Yuxuan, ZHENG Xiaoli, ZHANG Chuang

Computer Engineering and Applications 2023, 59 (13): 247-258. DOI: 10.3778/j.issn.1002-8331.2206-0190

Abstract （83）

PDF（pc）（826KB）（82）

Save

With the rapid iterative development of information technology, the problem of information overload is becoming more and more serious. The recommendation algorithm can solve the information overload to a certain extent, but the traditional recommendation algorithm can not effectively solve the related problems such as data sparsity and recommendation accuracy. This paper proposes a graph convolution attention collaborative filtering（GCACF） recommendation method. Firstly, the model obtains the relevant interactive information of users and projects and transforms into corresponding feature vectors. Secondly, the feature vector aggregates with the propagation of graph convolution neural network and the attention mechanism redistributes the aggregated weight coefficients. Finally, the BPR loss function optimizes aggregated eigenvector and the model obtains the final recommendation result. Through the comparative experiments on Movielens-1M and Amazon-baby on two public datasets, GCACF is superior to the baseline method in precision, recall, Mrr, hit and NDCG.

Reference | Related Articles | Metrics

Select

Execution Plan Selection for Parallel Queries Using Graph Neural Networks

TAO Wenxia, NIU Baoning, LIU Haonan

Computer Engineering and Applications 2023, 59 (13): 259-265. DOI: 10.3778/j.issn.1002-8331.2208-0300

Abstract （70）

PDF（pc）（521KB）（39）

Save

Queries constitute the largest proportion of workload of database systems（DBS）, and their efficiency affects the performance of DBS. The execution of a query is affected by other parallel queries, resulting in query interaction（QI）, which is the main factor that makes it difficult for query optimizers to select a good execution plan for parallel queries. An encoding scheme called features of plans based on operator（FPO） is proposed to represent execution plans. QI is reflected by data sharing and resource competition between operators. The plan selection model based on graph（PSG） is proposed. PSG takes operators as nodes, operator features as node features, and relations between operators as edges to generate heterogeneous graphs as inputs of the model. Considering that there are many kinds of relations between operators with different functions, relational graph convolutional network（RGCN） is used to aggregate information, obtain a graph representation of a query mix, and extract its QI. Through fully connected layers（FC）, an execution plan is selected for a query. The average accuracy of PSG is 47.3?percentage points higher than that of query optimizers in PostgreSQL.

Reference | Related Articles | Metrics

Select

Multi-Relationship Recommendation Model Based on User Interest-Aware

HU Xinrong, DENG Jiewen, LUO Ruiqi, LIU Junping, ZHU Qiang, PENG Tao

Computer Engineering and Applications 2023, 59 (11): 231-240. DOI: 10.3778/j.issn.1002-8331.2206-0148

Abstract （83）

PDF（pc）（782KB）（51）

Save

Because graph convolution network（GCN） can use the cooperation signals of high-order neighbors to learn the embedding of users and items better, it has been widely used in recommendation systems. However, in the current multi-relationship recommendation model based on GCN, the embedding learning of user nodes will be interfered by high-order adjacent users with dissimilar interests, so that users with different interests will get similar embedding after multi-layer graph convolution, resulting in the over-smoothing problem. Therefore, this paper proposes a multi-relationship recommendation model based on user interest-aware（IMRRM） because the above problem. First, the model uses a light graph convolutional network in the user-item heterogeneous interaction graph to obtain the graph structure information of each user. And then the subgraph generation module uses the user’s graph structure information and initial features to effectively identify users with similar interests, and group similar users and their interaction items into a subgraph. Finally, more accurate user embeddings are obtained by performing deep embedding learning in subgraphs to prevent unrelated high-order neighbors from propagating more negative information. Therefore, the IMRRM model reduces the influence of noise information on user embedding learning, effectively alleviates the over-smoothing problem, and makes multi-relationship recommendation more accurately. The effectiveness and robustness of IMRRM are verified by experiments on two public datasets, Beibei and Taobao. The experimental results show that the IMRRM model is improved by 1.98% and 1.49% on HR10, and 1.58% and 1.81% on NDCG10, respectively, with better performance.

Reference | Related Articles | Metrics

Select

Article Topic-Based Heterogeneous Information Network for Venue Recommendation

WANG Bingyuan, LIU Baisong, ZHANG Xueyuan, QIN Jiangcheng, DONG Qian, QIAN Jiangbo

Computer Engineering and Applications 2023, 59 (11): 241-250. DOI: 10.3778/j.issn.1002-8331.2207-0502

Abstract （68）

PDF（pc）（749KB）（43）

Save

With the expansion of academic information, scholars face a big challenge of efficiently selecting valid information in the era of big academic data. A venue recommendation system is one of the main ways to assist scholars in solving the information overload problem. This paper focuses on the issue of how to fit appropriate academic journals for manuscripts efficiently. It extracts diverse academic entities and edges from academic data for constructing academic heterogeneous information networks. This paper proposes a novel method for venue recommendation（SCVR）. Firstly, the topic information is extracted from the abstracts and topics by LDA and guides different types of nodes to map to the multi-topic feature space. Then, the meta-path contextual information is aggregated to the target node, forming a multi-topic node representation. Finally, the node representations from multiple mate-paths are combined into the final multi-topic node representations. SCVR learns the multi-topic node representations with paper content and network structure to venue recommendation.Experiments on two real academic datasets show that a heterogeneous information network recommendation incorporating article topics can effectively improve the performance of the venue recommendation. Compared with the current heterogeneous information network recommendation and traditional venue recommendation, the performance of SCVR has improved by an average of 2.7% and 19%, which indicates that SCVR has better performance in the area of venue recommendation.

Reference | Related Articles | Metrics

Select

Anomaly Series Detection Algorithm Based on Segmentation Feature Representation

SONG Chunlei, ZHAO Xujun, GAO Yaxing, JIN Guangyin

Computer Engineering and Applications 2023, 59 (9): 262-271. DOI: 10.3778/j.issn.1002-8331.2205-0271

Abstract （87）

PDF（pc）（820KB）（83）

Save

The supervised anomaly detection method of time series usually depends on the label of data, which not only consumes a lot of time for data labeling, but also is difficult to apply to data sets that cannot be given a label. In order to solve the labeling problem in anomaly series detection, an anomaly series detection algorithm based on segmentation feature representation is proposed. This method uses the idea of piecewise aggregation to standardize the calculation of time series, and obtains the characteristic representation of time series data, which can improve the reliability of anomaly detection of unlabeled time series. The expressed features are divided into abnormal series related features and irrelevant features. Pruning abnormal series irrelevant features can reduce the adverse impact of these features on the detection results. In order to effectively quantify the differences between different series, a time series similarity measurement method for time weight analysis is proposed, and the similarity matrix of time series is constructed to calculate the similarity between series, which can be applied to unlabeled time series. On this basis, the anomaly score of each sub-series is calculated according to the similarity matrix, which is used to determine the abnormal sub-series. Finally, the experimental comparison between synthetic data sets and real data sets shows that this method saves the computational overhead, improves the time efficiency of the algorithm and the accuracy of anomaly series detection.The supervised anomaly detection method of time series usually depends on the label of data, which not only consumes a lot of time for data labeling, but also is difficult to apply to data sets that cannot be given a label. In order to solve the labeling problem in anomaly series detection, an anomaly series detection algorithm based on segmentation feature representation is proposed. This method uses the idea of piecewise aggregation to standardize the calculation of time series, and obtains the characteristic representation of time series data, which can improve the reliability of anomaly detection of unlabeled time series. The expressed features are divided into abnormal series related features and irrelevant features. Pruning abnormal series irrelevant features can reduce the adverse impact of these features on the detection results. In order to effectively quantify the differences between different series, a time series similarity measurement method for time weight analysis is proposed, and the similarity matrix of time series is constructed to calculate the similarity between series, which can be applied to unlabeled time series. On this basis, the anomaly score of each sub-series is calculated according to the similarity matrix, which is used to determine the abnormal sub-series. Finally, the experimental comparison between synthetic data sets and real data sets shows that this method saves the computational overhead, improves the time efficiency of the algorithm and the accuracy of anomaly series detection.

Reference | Related Articles | Metrics

Select

Research on Group Preference Fusion Strategy Based on Two-Layer Attention Mechanism

MEI Yuzhu, HU Zhulin, ZHU Xinjuan

Computer Engineering and Applications 2023, 59 (9): 272-279. DOI: 10.3778/j.issn.1002-8331.2205-0437

Abstract （136）

PDF（pc）（532KB）（76）

Save

One of the current research hotspots of recommender systems and its evolution trend is that personalized recommendation gradually shifts from focusing on individual recommendation to focusing on group recommendation. At present, most group recommendation methods are accustomed to adopting a predefined static strategy when choosing a preference fusion strategy, and the characteristics of the static strategy make the algorithm unable to maximize the simulation of the real process of group decision-making. On the basis of previous research, this paper proposes a group recommendation method based on a two-layer attention mechanism, which fully takes into account the differences and mutual influence of group users, as well as the decision-making power in different fields. The attention weight of each member in the group is calculated to other members, the group member feature vector is obtained, and then the attention weight of each member in selecting a certain item is calculated, and the preference vector for the item for the group is generated. The interaction between group users and the process of group decision-making are fully restored. By comparing the CAMRa2011 and Meetup datasets with COM, SIG, AGR, AGREE, FastGR and other methods under different parameter conditions, the two indicators of normalized discount cumulative gain and hit rate are higher than the baseline model, is up to 0.025 4 and 0.030 7.

Reference | Related Articles | Metrics

Select

Traffic Flow Forecasting Model for Improved Spatio-Temporal Transformer

GAO Rong, WAN Yiliang, SHAO Xiongkai, Wu Xinyun

Computer Engineering and Applications 2023, 59 (7): 250-260. DOI: 10.3778/j.issn.1002-8331.2203-0290

Abstract （230）

PDF（pc）（844KB）（122）

Save

To address the low performance problem of traffic flow prediction model based on spatio-temporal Transformer model, an improved spatio-temporal Transformer model（ISTTM） based on encoder-decoder is proposed. The encoder encodes the historical traffic features and the decoder predicts the future sequences. Firstly, the encoder combines spatial sparse self-attentiveness and temporal hierarchical diffusion convolution to capture the dynamic spatial correlation and local spatial features of traffic flows, and then uses temporal self-attentiveness to model the nonlinear temporal correlation. Then, the decoder mines the spatio-temporal features of the input sequences similarly to the encoder. Finally, based on the spatio-temporal features extracted by the encoder-decoder, the impact of historical traffic observations on future forecasts is simulated using double cross-attention, modeling the direct relationship between each historical time step and each future time step and the impact on the whole future time period, and the final representation of the future traffic flow is output. To confirm the effectiveness of ISTTM, experiments are executed on two real-world large-scale datasets, METR-LA and NE-BJ, and the ISTTM results outperform the six state-of-the-art baselines.

Reference | Related Articles | Metrics

Select

Application of CEEMDAN-HURST Algorithm in COVID-19 Prediction

WANG Qiyun, ZHENG Zhongtuan

Computer Engineering and Applications 2023, 59 (7): 261-268. DOI: 10.3778/j.issn.1002-8331.2205-0253

Abstract （84）

PDF（pc）（696KB）（30）

Save

Considering the new COVID-19 cases are a nonlinear and non-stationary time series, a combined COVID-19 prediction model based on CEEMDAN-HURST algorithm is proposed. Firstly, the time series of newly confirmed cases are decomposed into sub-series with different frequencies using the complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN） algorithm. Secondly, the randomness of each sub-sequence is analyzed by HURST index and the sub-sequence is integrated into three sub-sequences of high frequency, medium frequency and low frequency. The three sub-sequences are predicted by the least square support vector machine（LSSVM）. Finally, the prediction results of each reconstructed subsequence are superimposed to obtain the final predicted value of newly confirmed COVID-19 cases. The results show that the COVID-19 new case combination prediction model based on CEEMDAN-HURST algorithm improves the efficiency and prediction accuracy in the nonlinear time series prediction process. Compared with the CEEMDAN-PE combined model, the mean absolute error and root mean square error are reduced by 11.13% and 29.67%, respectively, indicating that the CEEMDAN-HURST algorithm can effectively solve the problems of low prediction efficiency and low prediction accuracy commonly existed in nonlinear time series forecasting models. Meanwhile, HURST index measures the deviation degree of time series, and the HURST index is introduced to merge, reconstruct and integrate, which can reduce the number of sub-series needed for time series prediction.

Reference | Related Articles | Metrics

Select

Parallel OPTICS by Using Mean Distance and Relevance Marks

ZHENG Jian, YU Xin

Computer Engineering and Applications 2023, 59 (5): 232-244. DOI: 10.3778/j.issn.1002-8331.2203-0018

Abstract （67）

PDF（pc）（901KB）（39）

Save

The main target of this paper is to design a parallel optics algorithm by using mean distance and relevance marks based on MapReduce, noted as POMDRM-MR, to deal with the problems of unreasonable data division, low accuracy of clustering results, the results are greatly affected by parameters and low efficiency of parallelization in parallel density-based clustering algorithm in big data. In POMDRM-MR, an approach called partition with reduced boundary points based on dimension sparsity（DS-PRBP） is proposed to divide the dataset. For each partition, the algorithm called marking and ordering points to identify the cluster structure（MOPTICS） is proposed to construct the correlations between data points and core points and mark the number of iterations, the field mean distance strategy（FMD） is proposed to calculate the field mean distance of data points instead of the reachable distance in measuring distance. After outputting sequence, combined with reordering and extracting clusters algorithm（REC）, the sequence is sorted twice which improves the accuracy and stability. In merging global clusters, an approach called using boundary density to filter local cluster（BD-FLC） is used to calculate and filter local clusters with similar density. And based on the union-type merging of n-ary trees and MapReduce, the parallel local cluster merging algorithm（MCNT-MR） is proposed to get the clustering results faster and merge local clusters in parallel which improves efficiency of merging local clusters. The experiments show that POMDRM-MR algorithm has better effect, and better parallelization performance on large-scale datasets.

Reference | Related Articles | Metrics

Select

Hierarchical Grid Division to Realize Cluster and Scatter Visualization of Massive Map Markers

FU Chen’en, CHEN Qiong

Computer Engineering and Applications 2023, 59 (5): 245-251. DOI: 10.3778/j.issn.1002-8331.2203-0123

Abstract （72）

PDF（pc）（688KB）（32）

Save

In traditional map visualization, point clustering is used in the display of massive map markers, but all kinds of point clustering algorithms are run-time calculations without hierarchical mechanism, and there is no filtering mechanism for map markers stack when a large number of points are scattered and displayed. In response to this problem, a solution of clustering and scattering of massive map markers is proposed based on hierarchical grid division. This method builds a K-D tree index for the center point of the hierarchical grid, and builds a quadtree index for the massive points. Through the index and storage technology, the efficient query of clustering is realized. Add grid filtering to eliminate stacking problems when massive points are scattered. The comparison is carried out on the experimental case data set, and the results show that, compared with the traditional point clustering scheme, the computing performance is significantly improved in the case of a large amount of data, and the filtering algorithm is added to the scattered display of massive markers, which effectively improves the user experience.

Reference | Related Articles | Metrics

Select

Group Recommendation Method Integrating Probability Matrix Decomposition and ER Rules

WANG Yonggui, ZHANG Jian

Computer Engineering and Applications 2023, 59 (5): 252-261. DOI: 10.3778/j.issn.1002-8331.2203-0125

Abstract （69）

PDF（pc）（595KB）（48）

Save

Group recommendation needs to consider the preferences of all members in a group at the same time, and recommend items to the group by integrating member preferences. Most of the existing researches on group recommendation methods assign the same weight to all users in the group, without considering the importance and reliability of different group members in real life. To solve this problem, a group recommendation method（FPMF-ER） based on probabilistic matrix decomposition and evidence reasoning rules is proposed to improve the process of individual prediction and preference fusion in group recommendation. Firstly, the classical probability matrix decomposition is improved by combining user relationship information to obtain a more complete and accurate personal prediction score. Subsequently, ER rule is introduced in the process of preference fusion of group members, and the influence of group members is identified according to the weight and reliability of group members, so that preference fusion is more reasonable and accurate. In order to verify the effectiveness of the method, a comparative experiment is conducted on the Book-Crossing dataset. The experimental results show that compared with the optimal benchmark model, the accuracy of FPMF-ER recommendation results and user satisfaction are increased by at least 2.55% and 2.06%, respectively.

Reference | Related Articles | Metrics

Select

Node Importance Analysis Integrated with Community Assessment

SUN Baibing, SUN Jiazheng, HE Quan, DU Yanhui

Computer Engineering and Applications 2023, 59 (3): 226-233. DOI: 10.3778/j.issn.1002-8331.2205-0509

Abstract （97）

PDF（pc）（701KB）（52）

Save

The research on mining key members in the target is an important branch in the field of social networks, but existing importance algorithms are prone to the phenomenon of aggregation of mined key nodes. To address this problem, a node importance algorithm integrated with community assessment is proposed, which defines a community importance assessment function based on the network topology of the target group, incorporating the internal influence of members in their community and external connectivity to comprehensively evaluate the importance of members. Taking four different complex networks as experimental data, compared with the existing algorithms, it is verified from three dimensions：propagation ability, robustness and Kendall correlation coefficient. The experiments show that the algorithm is more accurate in measuring the importance of members in the group.

Reference | Related Articles | Metrics

Select

Construction Method of SDN for “Virtual-Real Fusion” Network Under Cloud

CHEN Yu, HAN Jiujiang, LIU Jian, XIAN Ming, WANG Huimei, ZHANG Yuxiang

Computer Engineering and Applications 2023, 59 (3): 234-243. DOI: 10.3778/j.issn.1002-8331.2204-0465

Abstract （126）

PDF（pc）（4559KB）（71）

Save

In the scenario construction of network simulation test, it is difficult for some complicated physical terminals to realize virtual simulation, so it is essential for some physical terminals to be connected. However, the existing access modes are not applicable to large-scale network scenarios due to complex deployment and network performance bottlenecks. In order to effectively solve these problems, this paper proposes a construction method of virtual-real fusion network simulation test based on SDN（software-defined-network）. SDN controller combined with OpenFlow construction algorithm is used to realize the link management and data connectivity of virtual and real networks. It designs a prototype system for virtual-real fusion network simulation. The SDN controller is used to realize the common networking of virtual instances in the cloud environment and the physical terminals outside the cloud, so as to build a network simulation with “combination of virtual-real”. The experiments show that the network simulation method based on SDN can unite the virtual equipment in the cloud and the real terminals outside the cloud to make large-scale network efficiently, and do well in network performance.

Reference | Related Articles | Metrics

Select

Self-Supervised Graph Neural Networks for Session-Based Recommendation

WANG Yonggui, ZHAO Xiaoxuan

Computer Engineering and Applications 2023, 59 (3): 244-252. DOI: 10.3778/j.issn.1002-8331.2204-0245

Abstract （138）

PDF（pc）（600KB）（72）

Save

Compared with traditional recommendation methods, session-based recommendation methods are more susceptible to the data sparsity problem due to the limited data of short-term user interaction. In order to enhance session data and alleviate the impact of datas parseness on session recommendation performance, a model named self-supervised graph neural networks for session-based recommendation（Ss-GNN） is proposed. It constructs asession graph and establishes a session recommendation task based on graph attention network to obtain the item-level and session-level representation. Then, from the perspective of the session-level representation, it uses the user’s general and current interests to construct an auxiliary task to obtain self-supervised signals.It utilizes self-supervised learning to maximize mutual information between the recommendation task and the auxiliary task to enhance session data and improve recommendation performance. Experiments are conducted on two public data sets, Yoochoose and Tmall. Compared with the baseline model, the proposed model improves at least 0.94% and 0.79% in P@20 and MRR@20 on Yoochoose, 9.61% and 4.67% in P@20 and MRR@20 on Tmall, which proves the effectiveness of Ss-GNN model.

Reference | Related Articles | Metrics

Select

A-DDPG：Research on Offloading of Multi-User Edge Computing System

CAO Shaohua, JIANG Jiajia, CHEN Shu, ZHAN Zijun, ZHANG Weishan

Computer Engineering and Applications 2023, 59 (1): 259-268. DOI: 10.3778/j.issn.1002-8331.2203-0461

Abstract （218）

PDF（pc）（1276KB）（146）

Save

In order to reduce the total cost of users in multi-user systems with multiple edge servers, a deep reinforcement learning offloading algorithm（A-DDPG） based on DDPG is proposed by combining deep deterministic policy gradient（DDPG）, long short term memory（LSTM） and attention mechanism, which uses binary offloading strategy and takes into account the latency sensitivity of tasks and the limited server load as well as task migration to adapt offload tasks to minimize the total loss caused by latency-sensitive task timeouts. Two metrics, latency and energy consumption, are considered and different weight values are set to address the unfairness caused by different user types, and the task offloading problem is formulated to minimize the total cost of all task completion latency and energy consumption, with the selection of target servers and the amount of data offloaded as learning objectives. The experimental results show that the A-DDPG algorithm has good stability and convergence, and the total user cost of the A-DDPG algorithm is reduced by 27% and 26.66% compared to the DDPG algorithm and the twin delayed deep deterministic policy gradient（TD3） algorithm respectively. It achieves better results in terms of reward, total cost and task failure rate, as the average time to reach the optimal task failure rate is 57.14% and 40% earlier, respectively.

Reference | Related Articles | Metrics

Select

Decentralized Federated Learning Strategy for Non-Independent and Identically Distributed Data

TAN Rongjie, HONG Zhiyong, YU Wenhua, ZENG Zhiqiang

Computer Engineering and Applications 2023, 59 (1): 269-277. DOI: 10.3778/j.issn.1002-8331.2204-0165

Abstract （256）

PDF（pc）（4322KB）（185）

Save

Blockchain has the characteristics of immutability and decentralization, and its combination with federated learning has become a hot topic in the field of artificial intelligence. At present, decentralized federated learning has the problem of performance degradation caused by non-IID distribution of training data. In order to solve this problem, the paper firstly proposes a calculation method of model similarity, and then designs a decentralized federated learning strategy based on the similarity of the model. The learning strategy is tested using five federated learning tasks, using the CNN model to train the fashion-mnist dataset, the alexnet model to train the cifar10 dataset, the TextRnn model to train the THUsnews dataset, the Resnet18 model to train the SVHN dataset and the LSTM model to train the sentiment140 dataset. The experimental results show that the designed strategy performs decentralized federated learning under the non-IID data of the five tasks, and the accuracy rates are increased by 2.51, 5.16, 17.58, 2.46 and 5.23?percentage points respectively.

Reference | Related Articles | Metrics

Select

Path Discrimination Method for Service Fault Propagation in Cloud Computing Environment

SI Jianzhe, JIANG Ying, LI Rongchen, CHEN Weiwei

Computer Engineering and Applications 2022, 58 (23): 94-103. DOI: 10.3778/j.issn.1002-8331.2203-0239

Abstract （84）

PDF（pc）（1068KB）（46）

Save

Since cloud computing services can be characterized by dynamic, complexity and diversity, it makes the interaction between different services more complicated. As service failures happen frequently in cloud computing environment and faults are constantly spreading, resulting in system failure. Most of the existing fault propagation analysis methods all have some problems, such as over reliance on the historical data, only measuring single fault propagation factor, and failing to model the dynamic change system structure. In order to solve these problems, a method of service path discrimination method for service fault propagation in cloud computing environment is proposed. Firstly, the service interaction diagram is dynamically established. Secondly, the structure of service interaction diagram is optimized and service relationship diagram is set up. Thirdly, the possibility of service failure is calculated by considering service operating and environmental conditions, and the fault service is determined. Finally, the influencing factors of service fault propagation are comprehensively analyzed, and the service fault propagation probability is calculated, and the path discrimination of service fault propagation is carried out. Experimental results show that this method can accurately determine the fault service and effectively distinguish the service fault propagation path.

Reference | Related Articles | Metrics

Select

Top-k Collective Spatial Keyword Approximation Query

MENG Xiangfu, WANG Dandan, ZHANG Xiaoyan, JIA Jianghao

Computer Engineering and Applications 2022, 58 (23): 104-116. DOI: 10.3778/j.issn.1002-8331.2201-0393

Abstract （83）

PDF（pc）（979KB）（33）

Save

In recent years, the scale of spatio-textual data with location and text information has been increasing rapidly, and the spatial keyword query technology against spatio-textual data has been widely studied and applied. Most of the existing spatial keyword query methods usually take a single spatial object as the basic unit of the query results. Recently, there are a few recent research work aiming to find a group of spatial object as the basic unit of the query results, this group of spatial objects jointly meet the requirements of the given spatial keyword query. However, such kind of methods does not consider the relationships（such as social correlations， textual similarity） between spatial objects in the group. To deal with this problem, this paper proposes a top-[k] collective spatial keyword approximate query method. First, an association rule-based social relationship evaluation method for spatial objects is proposed. Furthermore, it designs a scoring function which combines the location distances and social relationships of spatial objects within a group. Second, a VP-Tree based pruning strategy is proposed for quickly searching the local neighborhood of spatial objects. Last, the top-[k] spatial object groups are selected as the query result by leveraging the scoring function to calculate the score of candidate spatial object groups. The experimental results show that the proposed spatial object social relationship evaluation method can achieve high accuracy, the proposed pruning strategy has high execution efficiency, and the obtained top-[k]groups of spatial objects can meet user need and preferences well.

Reference | Related Articles | Metrics

Select

Collaborative Computing Offloading Strategy in Mobile Edge Computing

LI Shun, GE Haibo, LIU Linhuan, CHEN Xutao

Computer Engineering and Applications 2022, 58 (21): 83-90. DOI: 10.3778/j.issn.1002-8331.2201-0015

Abstract （155）

PDF（pc）（713KB）（113）

Save

Aiming at the problem of resource waste in the idle state of remote edge servers when a single edge server is uninstalled, this paper proposes an edge cloud collaborative computing offloading strategy based on an improved hybrid particle swarm algorithm（cross reorganization PSO, CRPSO）. In this offloading strategy, a model is established with the goal of minimizing the total cost of the system（the weighted sum of time delay and energy consumption）. In the particle swarm algorithm, the fitness is used to group the advantages and disadvantages of the particles, and the disadvantages are solved by introducing the crossover idea in the genetic algorithm. The group of particles is optimized, the particles in the original population are optimized by a two-layer screening mechanism, and the optimal unloading strategy of the task is achieved through algorithm iteration. The simulation results show that, compared with the Local-MEC algorithm, ECPSO algorithm and GCPSO algorithm, the proposed CRPSO algorithm has the smallest total system cost and the optimization effect is obvious.

Reference | Related Articles | Metrics

Select

Exploring Message Passing Method Between Spark Tasks

XIA Libin, LIU Xiaoyu, SUN Wei, JIANG Xiaowei, SUN Gongxing

Computer Engineering and Applications 2022, 58 (21): 91-97. DOI: 10.3778/j.issn.1002-8331.2112-0182

Abstract （123）

PDF（pc）（1201KB）（41）

Save

Engineering problems and scientific research are facing dual challenges of big data processing and high-performance computing tasks. Spark, a distributed processing framework based on in-memory computing technology, has been widely used in academia and industry. However, its MapReduce-like programming model fails to communicate between tasks, causing numerical algorithms in scientific computing cannot be efficiently implemented. In response to the above problems, a computing system is proposed in this paper that combines Spark in-memory computing model with MPI message passing, which takes full advantage of the fast speed of memory access and multiple high performance communication mechanisms of MPI. It can not only supplement the insufficient expressiveness of the Spark programming model, but also provide a data-oriented DAG computation method for MPI. Internal runtime environment and scheduling strategy of Spark are modified to seamlessly integrate MPI into Spark to provide a unified in-memory computing system for high-performance computing and big data processing tasks. The tests indicate that the performance of numerical computation and iterative algorithm is improved by at least 50% compared with Spark.

Reference | Related Articles | Metrics

Select

Research on Service Quality Model for SDN-Enabled Cloud Computing Platform

ZHENG Bing, LI Hua

Computer Engineering and Applications 2022, 58 (21): 98-108. DOI: 10.3778/j.issn.1002-8331.2111-0525

Abstract （90）

PDF（pc）（695KB）（45）

Save

The SDN-enabled cloud computing platform has gradually been used widely, and more and more attention has been paid to the quality evaluation of services provided by SDN-enabled cloud computing platform. Firstly, the general cloud service and the software-based characteristics of SDN-enabled cloud computing platform services are considered. From the perspectives of service providers and users, the design requirements of service quality model are analyzed, and the categorical dimensions of service quality model are obtained. Then, referring to SQuaRE international standard, the service quality model QM-SDNCCP of SDN-enabled cloud computing platform is established, and the calculation method of the metrics is given. Furthermore, the purpose of describing and quantifying the service quality of SDN-enabled cloud computing platform is achieved. Finally, the reliability analysis of QM-SDNCCP is carried out, which shows that the design of QM-SDNCCP model characteristics and sub-characteristics are reliable and consistent, it means QM-SDNCCP can be applied in the stages of technology selection, acceptance, operation and service management and also it can be used to guide on selection of cloud platform service and quality improvement by service providers and users.

Reference | Related Articles | Metrics

Select

Recommendation Algorithm Integrating Collaborative Knowledge Graph and Optimizing Graph Attention Network

TANG Hong, FAN Sen, TANG Fan

Computer Engineering and Applications 2022, 58 (19): 98-106. DOI: 10.3778/j.issn.1002-8331.2201-0183

Abstract （220）

PDF（pc）（809KB）（78）

Save

Considering the problems of data sparseness and high model complexity in recommendation algorithms, this paper proposes a recommendation model that integrates collaborative knowledge graphs and optimized graph attention networks. Firstly, the knowledge graph and the user-item interaction graph are combined into a collaborative knowledge graph and embedded into the optimized graph attention network model, which cannot only alleviate the data sparsity problem well, but also mine potential interests and high order relationship of users. Secondly, using an optimized graph convolutional network, by removing feature transformation and nonlinear activation modules, the model complexity can be greatly reduced without affecting the overall recommendation performance. Combined with the deviation-based attention mechanism, the deviation between the candidate item and the user’s real interest item can be sensed in time to improve the training efficiency of the model. Finally, simulation experiments are carried out on the Movielens dataset and the Double dataset, and it is concluded that the recommended performance and time complexity of the algorithm are effectively improved compared with the comparison algorithm.

Reference | Related Articles | Metrics

Select

Financial Time Series Forecasting Method Based on Multi-Scale Features and Attention Mechanism

ZHAN Xi, PAN Zhisong, LI Wei, ZHANG Yanyan, BAI Wei WANG Cailing

Computer Engineering and Applications 2022, 58 (19): 107-115. DOI: 10.3778/j.issn.1002-8331.2203-0138

Abstract （188）

PDF（pc）（1099KB）（81）

Save

Financial time series forecasting is a very important practical problem in the economic field. However, due to the noise and volatility of financial markets, the forecasting accuracy of current existing methods is not yet satisfactory. In order to improve the prediction accuracy of financial time series, a hybrid prediction model DCNN_LSTM_AT integrating dilated convolutional neural network（DCNN）, long short term memory（LSTM） and attention mechanism（AT） is proposed. The model consists of two parts：the first part contains a dilated convolutional neural network and an LSTM-based encoder, whose function is to extract effective information at different temporal scales in the original sequence data. The second part is composed of an LSTM decoder with an attention mechanism, its function is to filter the information extracted from the first part and use the filtered information to make predictions. Finally, the proposed model is tested on 3 index datasets and 3 individual stock datasets, and compared with other common benchmark models. The experimental results show that the model has better prediction accuracy and stability than other models.

Reference | Related Articles | Metrics

Select

Performance Optimization of Blockchain Sharding System Combined with Deep Reinforcement Learning

WEN Jianwei, YAO Bingbing, WAN Jianxiong, LI Leixiao

Computer Engineering and Applications 2022, 58 (19): 116-123. DOI: 10.3778/j.issn.1002-8331.2203-0142

Abstract （196）

PDF（pc）（613KB）（96）

Save

Improving the throughput of the blockchain system is one of the key issues for the widespread application of blockchain. In view of the above problems, the sharding technology is applied to the blockchain system, and the throughput of the blockchain is improved by making the blockchain process transactions in parallel. The blockchain shard selection problem is established as a Markov decision process（MDP）, and a blockchain shard optimal selection strategy（BDQSB） based on deep reinforcement learning（DRL） is designed. The adopted BDQSB algorithm overcomes the shortcomings of the traditional DRL algorithm with high behavior space dimension and slow neural network training. The simulation results show that the proposed method can effectively reduce the behavior space dimension and improve the throughput and scalability of blockchain processing transactions.

Reference | Related Articles | Metrics

Select

Patent Knowledge Recommendation Service Based on Deep Learning

LI Zhenyu, ZHAN Hongfei, YU Junhe, WANG Rui, DENG Huijun

Computer Engineering and Applications 2022, 58 (15): 95-109. DOI: 10.3778/j.issn.1002-8331.2107-0418

Abstract （153）

PDF（pc）（2052KB）（92）

Save

Patent contains the most complete design information in most domains, it can provide designers with valuable guidance in solving design problems. Aiming at the problem that the existing patent recommendation methods are difficult to effectively recommend cross-domain patents, a cross-domain patent knowledge recommendation method based on deep learning is proposed for the conceptual design of innovative products. The product function and knowledge demand situation are modeled, the design problem is standardized, and the design problem space is generated. A semi-supervised learning algorithm（TG-TCI） is proposed to automatically classify and label patent function information according to the function base, and use entity recognition algorithm（BERT-BiLSTM-CRF） to extract patent application scenario terms and technical terms, combined with international patent classifications（IPC） information is used to represent the function, context, technology and domain attributes of patents, thereby generating patent knowledge space. Find the required cross-domain patents through the functional base and knowledge context mapping from the design problem space to the patent knowledge space, cluster and evaluate them according to the technology and domain attributes, and select specific patents to stimulate the creativity of designers. An actual case is used to analyze and verify the feasibility and effectiveness of the patent knowledge recommendation model based on deep learning.

Reference | Related Articles | Metrics

Select

Deep Convolutional Neural Network Algorithm Based on Feature Map in Big Data Environment

MAO Yimin, ZHANG Ruipeng, GAO Bo

Computer Engineering and Applications 2022, 58 (15): 110-116. DOI: 10.3778/j.issn.1002-8331.2010-0081

Abstract （144）

PDF（pc）（709KB）（43）

Save

Aiming at problems such as excessive network redundant parameters, poor parameter optimization ability and low parallel efficiency exist in DCNN（deep convolutional neural network） algorithm under big data environment. In this paper, a deep convolutional neural network algorithm based on feature graph and parallel computational entropy is proposed. The algorithm is MR-FPDCNN（deep convolutional neural network algorithm based on feature graph and parallel computing entropy using MapRuduce）. The algorithm designs the FMPTL（feature map pruning based on Taylor loss） and the pre-training network to obtain the compressed DCNN, which effectively reduces the redundant parameters and also reduces the computational cost of DCNN training. This paper proposes the IFAS based on ISS, initializes DCNN parameters according to the “IFAS” algorithm, realizes the parallelization training of DCNN, and improves the optimization ability of network. In the Reduce phase, a DLBPCE（dynamic load balancing strategy based on parallel computing entropy） is proposed to obtain global training results, realizing fast uniform grouping of data and increasing the acceleration ratio of the parallel system. Experimental results show that this algorithm not only reduces the computational cost of DCNN training in big data environment, but also improves the parallelization performance of parallel system.

Reference | Related Articles | Metrics

Select

Two-Level Multi-Access Edge Computing Energy-Saving Offloading Strategy for IoT

DENG Yu, ZHAO Junhui, ZHANG Qingmiao

Computer Engineering and Applications 2022, 58 (13): 94-101. DOI: 10.3778/j.issn.1002-8331.2108-0069

Abstract （99）

PDF（pc）（924KB）（77）

Save

Multi-access edge computing（MEC） technology sinks computing and storage resources to the edge of the network, which can greatly improve the computing power and real-time performance of the Internet of things（IoT） system. However, MEC is often constrained by the growth of computing demand and energy constraints. Thus, efficient computing offloading and energy consumption optimization mechanism is an important research direction in MEC technology. To ensure the computing efficiency while maximizing the energy efficiency in the computing process, a two-level edge nodes（ENs） relay network model is proposed, and an optimal energy consumption algorithm（OECA） for joint optimization of computing resources and channel resources is designed. Firstly, the energy efficiency in MEC is modeled as a 0-1 knapsack problem. Secondly, the system adaptively selects the computing mode and allocates wireless channel resources, aiming at minimizing the overall energy consumption of the system. Finally, the OECA algorithm is simulated and verified in Python environment. The simulation results show that OECA can increase the network capacity by 18.3% and reduce the energy consumption by 13.1% compared with the offloading strategy algorithm based on directed acyclic graph algorithm（DAGA）.

Reference | Related Articles | Metrics

Content of Big Data and Cloud Computing in our journal