With the rapid development of Internet of Things(IoT)technology,IoT systems have been widely applied in health-care,transportation,home,and other fields.However,with the continuous expansion of the scale and increasin...With the rapid development of Internet of Things(IoT)technology,IoT systems have been widely applied in health-care,transportation,home,and other fields.However,with the continuous expansion of the scale and increasing complexity of IoT systems,the stability and security issues of IoT systems have become increasingly prominent.Thus,it is crucial to detect anomalies in the collected IoT time series from various sensors.Recently,deep learning models have been leveraged for IoT anomaly detection.However,owing to the challenges associated with data labeling,most IoT anomaly detection methods resort to unsupervised learning techniques.Nevertheless,the absence of accurate abnormal information in unsupervised learning methods limits their performance.To address these problems,we propose AS-GCN-MTM,an adaptive structural Graph Convolutional Networks(GCN)-based framework using a mean-teacher mechanism(AS-GCN-MTM)for anomaly identification.It performs better than unsupervised methods using only a small amount of labeled data.Mean Teachers is an effective semi-supervised learning method that utilizes unlabeled data for training to improve the generalization ability and performance of the model.However,the dependencies between data are often unknown in time series data.To solve this problem,we designed a graph structure adaptive learning layer based on neural networks,which can automatically learn the graph structure from time series data.It not only better captures the relationships between nodes but also enhances the model’s performance by augmenting key data.Experiments have demonstrated that our method improves the baseline model with the highest F1 value by 10.4%,36.1%,and 5.6%,respectively,on three real datasets with a 10%data labeling rate.展开更多
Active learning in semi-supervised classification involves introducing additional labels for unlabelled data to improve the accuracy of the underlying classifier.A challenge is to identify which points to label to bes...Active learning in semi-supervised classification involves introducing additional labels for unlabelled data to improve the accuracy of the underlying classifier.A challenge is to identify which points to label to best improve performance while limiting the number of new labels."Model Change"active learning quantifies the resulting change incurred in the classifier by introducing the additional label(s).We pair this idea with graph-based semi-supervised learning(SSL)methods,that use the spectrum of the graph Laplacian matrix,which can be truncated to avoid prohibitively large computational and storage costs.We consider a family of convex loss functions for which the acquisition function can be efficiently approximated using the Laplace approximation of the posterior distribution.We show a variety of multiclass examples that illustrate improved performance over prior state-of-art.展开更多
Graph data have extensive applications in various domains,including social networks,biological reaction networks,and molecular structures.Graph classification aims to predict the properties of entire graphs,playing a ...Graph data have extensive applications in various domains,including social networks,biological reaction networks,and molecular structures.Graph classification aims to predict the properties of entire graphs,playing a crucial role in many downstream applications.However,existing graph neural network methods require a large amount of labeled data during the training process.In real-world scenarios,the acquisition of labels is extremely costly,resulting in labeled samples typically accounting for only a small portion of all training data,which limits model performance.Current semi-supervised graph classification methods,such as those based on pseudo-labels and knowledge distillation,still face limitations in effectively utilizing unlabeled graph data and mitigating pseudo-label bias issues.To address these challenges,we propose a Semi-supervised graph Contrastive learning based on Associative Memory network and Pseudo-label Similarity(SCoAMPS).SCoAMPS integrates pseudo-labeling techniques with contrastive learning by generating contrastive views through multiple encoders,selecting positive and negative samples using pseudo-label similarity,and defining associative memory network to alleviate pseudo-label bias problems.Experimental results demonstrate that SCoAMPS achieves significant performance improvements on multiple public datasets.展开更多
Within any scientific disciplines, a large amount of data are buried within various literature depositories and archives, making it difficult to manually extract useful information from the datum swamps. The machine-l...Within any scientific disciplines, a large amount of data are buried within various literature depositories and archives, making it difficult to manually extract useful information from the datum swamps. The machine-learning extraction of data therefore is necessary for the big-data-based studies. Here, we develop a new text-mining technique to reconstruct the global database of the Precambrian to Recent stromatolites, providing better understanding of secular changes of stromatolites though geological time. The step-by-step data extraction process is described as below. First, the PDF documents of stromatolite-containing literatures were collected, and converted into text formation. Second, a glossary and tag-labeling system using NLP(Natural Language Processing) software was employed to search for all possible candidate pairs from each sentence within the papers collected here. Third, each candidate pair and features were represented as a factor graph model using a series of heuristic procedures to score the weights of each pair feature. Occurrence data of stromatolites versus stratigraphical units(abbreviated as Strata), facies types, locations, and age worldwide were extracted from literatures, respectively, and their extraction accuracies are 92%/464, 87%/778, 92%/846, and 93%/405 from 3 750 scientific abstracts, respectively, and are 90%/1 734, 86%/2 869, 90%/2 055 and 91%/857 from 11 932 papers, respectively. A total of 10 072 unique datum items were identified. The newly obtained stromatolite dataset demonstrates that their stratigraphical occurrences reached a pronounced peak during the Proterozoic(2 500 – 541 Ma), followed by a distinct fall during the Early Phanerozoic, and overall fluctuations through the Phanerozoic(541–0 Ma). Globally, seven stromatolite hotspots were identified from the new dataset, including western United States, eastern United States, western Europe, India, South Africa, northern China, and southern China. The proportional occurrences of inland aquatic stromatolites remain rather low(~20%) in comparison to marine stromatolites from the Precambrian to Jurassic, and then display a significant increase(30%–70%) from the Cretaceous to the present.展开更多
Lateral interaction in the biological brain is a key mechanism that underlies higher cognitive functions.Linear self‐organising map(SOM)introduces lateral interaction in a general form in which signals of any modalit...Lateral interaction in the biological brain is a key mechanism that underlies higher cognitive functions.Linear self‐organising map(SOM)introduces lateral interaction in a general form in which signals of any modality can be used.Some approaches directly incorporate SOM learning rules into neural networks,but incur complex operations and poor extendibility.The efficient way to implement lateral interaction in deep neural networks is not well established.The use of Laplacian Matrix‐based Smoothing(LS)regularisation is proposed for implementing lateral interaction in a concise form.The authors’derivation and experiments show that lateral interaction implemented by SOM model is a special case of LS‐regulated k‐means,and they both show the topology‐preserving capability.The authors also verify that LS‐regularisation can be used in conjunction with the end‐to‐end training paradigm in deep auto‐encoders.Additionally,the benefits of LS‐regularisation in relaxing the requirement of parameter initialisation in various models and improving the classification performance of prototype classifiers are evaluated.Furthermore,the topologically ordered structure introduced by LS‐regularisation in feature extractor can improve the generalisation performance on classification tasks.Overall,LS‐regularisation is an effective and efficient way to implement lateral interaction and can be easily extended to different models.展开更多
Graph-based semi-supervised learning is an important semi-supervised learning paradigm. Although graphbased semi-supervised learning methods have been shown to be helpful in various situations, they may adversely affe...Graph-based semi-supervised learning is an important semi-supervised learning paradigm. Although graphbased semi-supervised learning methods have been shown to be helpful in various situations, they may adversely affect performance when using unlabeled data. In this paper, we propose a new graph-based semi-supervised learning method based on instance selection in order to reduce the chances of performance degeneration. Our basic idea is that given a set of unlabeled instances, it is not the best approach to exploit all the unlabeled instances; instead, we should exploit the unlabeled instances that are highly likely to help improve the performance, while not taking into account the ones with high risk. We develop both transductive and inductive variants of our method. Experiments on a broad range of data sets show that the chances of performance degeneration of our proposed method are much smaller than those of many state-of-the-art graph-based semi-supervised learning methods.展开更多
Lithium-sulfur(Li-S)batteries are notable for their high theoretical energy density,but the‘shuttle effect’and the limited conversion kinetics of Li-S species can downgrade their actual performance.An essential stra...Lithium-sulfur(Li-S)batteries are notable for their high theoretical energy density,but the‘shuttle effect’and the limited conversion kinetics of Li-S species can downgrade their actual performance.An essential strategy is to design anchoring materials(AMs)to appropriately adsorb Li-S species.Herein,we propose a new three-procedure protocol,named InfoAd(Informative Adsorption)to evaluate the anchoring of Li_(2)S on two-dimensional(2D)materials and disclose the underlying importance of material features by combining high-throughput calculation workflow and machine learning(ML).In this paradigm,we calculate the anchoring of Li_(2)S on 12552D A_(x)B_(y)(B in the VIA/VIIA group)materials and pick out 44(un)reported nontoxic 2D binary A_(x)B_(y)AMs,in which the importance of the geometric features on the anchoring effect is revealed by ML for the first time.We develop a new Infograph model for crystals to accurately predict whether a material has a moderate binding with Li_(2)S and extend it to all 2D materials.Our InfoAd protocol elucidates the underlying structure-property relationship of Li_(2)S adsorption on 2D materials and provides a general research framework of adsorption-related materials for catalysis and energy/substance storage.展开更多
Recently,graph neural networks(GNNs)have achieved remarkable performance in representation learning on graph-structured data.However,as the number of network layers increases,GNNs based on the neighborhood aggregation...Recently,graph neural networks(GNNs)have achieved remarkable performance in representation learning on graph-structured data.However,as the number of network layers increases,GNNs based on the neighborhood aggregation strategy deteriorate due to the problem of oversmoothing,which is the major bottleneck for applying GNNs to real-world graphs.Many efforts have been made to improve the process of feature information aggregation from directly connected nodes,i.e.,breadth exploration.However,these models perform the best only in the case of three or fewer layers,and the performance drops rapidly for deep layers.To alleviate oversmoothing,we propose a nested graph attention network(NGAT),which can work in a semi-supervised manner.In addition to breadth exploration,a k-layer NGAT uses a layer-wise aggregation strategy guided by the attention mechanism to selectively leverage feature information from the k;-order neighborhood,i.e.,depth exploration.Even with a 10-layer or deeper architecture,NGAT can balance the need for preserving the locality(including root node features and the local structure)and aggregating the information from a large neighborhood.In a number of experiments on standard node classification tasks,NGAT outperforms other novel models and achieves state-of-the-art performance.展开更多
The data generated from non-Euclidean domains and its graphical representation(with complex-relationship object interdependence)applications has observed an exponential growth.The sophistication of graph data has pose...The data generated from non-Euclidean domains and its graphical representation(with complex-relationship object interdependence)applications has observed an exponential growth.The sophistication of graph data has posed consequential obstacles to the existing machine learning algorithms.In this study,we have considered a revamped version of a semi-supervised learning algorithm for graph-structured data to address the issue of expanding deep learning approaches to represent the graph data.Additionally,the quantum information theory has been applied through Graph Neural Networks(GNNs)to generate Riemannian metrics in closed-form of several graph layers.In further,to pre-process the adjacency matrix of graphs,a new formulation is established to incorporate high order proximities.The proposed scheme has shown outstanding improvements to overcome the deficiencies in Graph Convolutional Network(GCN),particularly,the information loss and imprecise information representation with acceptable computational overhead.Moreover,the proposed Quantum Graph Convolutional Network(QGCN)has significantly strengthened the GCN on semi-supervised node classification tasks.In parallel,it expands the generalization process with a significant difference by making small random perturbationsG of the graph during the training process.The evaluation results are provided on three benchmark datasets,including Citeseer,Cora,and PubMed,that distinctly delineate the superiority of the proposed model in terms of computational accuracy against state-of-the-art GCN and three other methods based on the same algorithms in the existing literature.展开更多
As an important method for knowledge graph(KG)complementation,link prediction has become a hot research topic in recent years.In this paper,a performance enhancement scheme for link prediction models based on the idea...As an important method for knowledge graph(KG)complementation,link prediction has become a hot research topic in recent years.In this paper,a performance enhancement scheme for link prediction models based on the idea of semi-supervised learning and model soup is proposed,which effectively improves the model performance on several mainstream link prediction models with small changes to their architecture.This novel scheme consists of two main parts,one is predicting potential fact triples in the graph with semi-supervised learning strategies,the other is creatively combining semi-supervised learning and model soup to further improve the final model performance without adding significant computational overhead.Experiments validate the effectiveness of the scheme for a variety of link prediction models,especially on the dataset with dense relationships.In terms of CompGCN,the model with the best overall performance among the tested models improves its Hits@1 metric by 14.7%on the FB15K-237 dataset and 7.8%on the WN18RR dataset after using the enhancement scheme.Meanwhile,it is observed that the semi-supervised learning strategy in the augmentation scheme has a significant improvement for multi-class link prediction models,and the performance improvement brought by the introduction of the model soup is related to the specific tested models,as the performances of some models are improved while others remain largely unaffected.展开更多
Due to the many types of distributed denial-of-service attacks(DDoS)attacks and the large amount of data generated,it becomes a chal-lenge to manage and apply the malicious behavior knowledge generated by DDoS attacks...Due to the many types of distributed denial-of-service attacks(DDoS)attacks and the large amount of data generated,it becomes a chal-lenge to manage and apply the malicious behavior knowledge generated by DDoS attacks.We propose a malicious behavior knowledge base framework for DDoS attacks,which completes the construction and application of a multi-domain malicious behavior knowledge base.First,we collected mali-cious behavior traffic generated by five mainstream DDoS attacks.At the same time,we completed the knowledge collection mechanism through data pre-processing and dataset design.Then,we designed a malicious behavior category graph and malicious behavior structure graph for the characteristic information and spatial structure of DDoS attacks and completed the knowl-edge learning mechanism using a graph neural network model.To protect the data privacy of multiple multi-domain malicious behavior knowledge bases,we implement the knowledge-sharing mechanism based on federated learning.Finally,we store the constructed knowledge graphs,graph neural network model,and Federated model into the malicious behavior knowledge base to complete the knowledge management mechanism.The experimental results show that our proposed system architecture can effectively construct and apply the malicious behavior knowledge base,and the detection capability of multiple DDoS attacks occurring in the network reaches above 0.95,while there exists a certain anti-interference capability for data poisoning cases.展开更多
The continuous accumulation of operational data has provided an ideal platform to devise and implement customized data analytics for smart HVAC fault detection and diagnosis.In practice,the potentials of advanced supe...The continuous accumulation of operational data has provided an ideal platform to devise and implement customized data analytics for smart HVAC fault detection and diagnosis.In practice,the potentials of advanced supervised learning algorithms have not been fully realized due to the lack of sufficient labeled data.To tackle such data challenges,this study proposes a graph neural network-based approach to effectively utilizing both labeled and unlabeled operational data for optimum decision-makings.More specifically,a graph generation method is proposed to transform tabular building operational data into association graphs,based on which graph convolutions are performed to derive useful insights for fault classifications.Data experiments have been designed to evaluate the values of the methods proposed.Three datasets on HVAC air-side operations have been used to ensure the generalizability of results obtained.Different data scenarios,which vary in training data amounts and imbalance ratios,have been created to comprehensively quantify behavioral patterns of representative graph convolution networks and their architectures.The research results indicate that graph neural networks can effectively leverage associations among labeled and unlabeled data samples to achieve an increase of 2.86%–7.30%in fault classification accuracies,providing a novel and promising solution for smart building management.展开更多
基金This research is partially supported by the National Natural Science Foundation of China under Grant No.62376043Science and Technology Program of Sichuan Province under Grant Nos.2020JDRC0067,2023JDRC0087,and 24NSFTD0025.
文摘With the rapid development of Internet of Things(IoT)technology,IoT systems have been widely applied in health-care,transportation,home,and other fields.However,with the continuous expansion of the scale and increasing complexity of IoT systems,the stability and security issues of IoT systems have become increasingly prominent.Thus,it is crucial to detect anomalies in the collected IoT time series from various sensors.Recently,deep learning models have been leveraged for IoT anomaly detection.However,owing to the challenges associated with data labeling,most IoT anomaly detection methods resort to unsupervised learning techniques.Nevertheless,the absence of accurate abnormal information in unsupervised learning methods limits their performance.To address these problems,we propose AS-GCN-MTM,an adaptive structural Graph Convolutional Networks(GCN)-based framework using a mean-teacher mechanism(AS-GCN-MTM)for anomaly identification.It performs better than unsupervised methods using only a small amount of labeled data.Mean Teachers is an effective semi-supervised learning method that utilizes unlabeled data for training to improve the generalization ability and performance of the model.However,the dependencies between data are often unknown in time series data.To solve this problem,we designed a graph structure adaptive learning layer based on neural networks,which can automatically learn the graph structure from time series data.It not only better captures the relationships between nodes but also enhances the model’s performance by augmenting key data.Experiments have demonstrated that our method improves the baseline model with the highest F1 value by 10.4%,36.1%,and 5.6%,respectively,on three real datasets with a 10%data labeling rate.
基金supported by the DOD National Defense Science and Engineering Graduate(NDSEG)Research Fellowshipsupported by the NGA under Contract No.HM04762110003.
文摘Active learning in semi-supervised classification involves introducing additional labels for unlabelled data to improve the accuracy of the underlying classifier.A challenge is to identify which points to label to best improve performance while limiting the number of new labels."Model Change"active learning quantifies the resulting change incurred in the classifier by introducing the additional label(s).We pair this idea with graph-based semi-supervised learning(SSL)methods,that use the spectrum of the graph Laplacian matrix,which can be truncated to avoid prohibitively large computational and storage costs.We consider a family of convex loss functions for which the acquisition function can be efficiently approximated using the Laplace approximation of the posterior distribution.We show a variety of multiclass examples that illustrate improved performance over prior state-of-art.
基金supported by the Science and Technology Project of State Grid Jiangsu Electric Power Co.Ltd.(No.J2023153).
文摘Graph data have extensive applications in various domains,including social networks,biological reaction networks,and molecular structures.Graph classification aims to predict the properties of entire graphs,playing a crucial role in many downstream applications.However,existing graph neural network methods require a large amount of labeled data during the training process.In real-world scenarios,the acquisition of labels is extremely costly,resulting in labeled samples typically accounting for only a small portion of all training data,which limits model performance.Current semi-supervised graph classification methods,such as those based on pseudo-labels and knowledge distillation,still face limitations in effectively utilizing unlabeled graph data and mitigating pseudo-label bias issues.To address these challenges,we propose a Semi-supervised graph Contrastive learning based on Associative Memory network and Pseudo-label Similarity(SCoAMPS).SCoAMPS integrates pseudo-labeling techniques with contrastive learning by generating contrastive views through multiple encoders,selecting positive and negative samples using pseudo-label similarity,and defining associative memory network to alleviate pseudo-label bias problems.Experimental results demonstrate that SCoAMPS achieves significant performance improvements on multiple public datasets.
基金supported by three grants from the National Natural Science Foundation of China (Nos.41821001,41902315,41930322)。
文摘Within any scientific disciplines, a large amount of data are buried within various literature depositories and archives, making it difficult to manually extract useful information from the datum swamps. The machine-learning extraction of data therefore is necessary for the big-data-based studies. Here, we develop a new text-mining technique to reconstruct the global database of the Precambrian to Recent stromatolites, providing better understanding of secular changes of stromatolites though geological time. The step-by-step data extraction process is described as below. First, the PDF documents of stromatolite-containing literatures were collected, and converted into text formation. Second, a glossary and tag-labeling system using NLP(Natural Language Processing) software was employed to search for all possible candidate pairs from each sentence within the papers collected here. Third, each candidate pair and features were represented as a factor graph model using a series of heuristic procedures to score the weights of each pair feature. Occurrence data of stromatolites versus stratigraphical units(abbreviated as Strata), facies types, locations, and age worldwide were extracted from literatures, respectively, and their extraction accuracies are 92%/464, 87%/778, 92%/846, and 93%/405 from 3 750 scientific abstracts, respectively, and are 90%/1 734, 86%/2 869, 90%/2 055 and 91%/857 from 11 932 papers, respectively. A total of 10 072 unique datum items were identified. The newly obtained stromatolite dataset demonstrates that their stratigraphical occurrences reached a pronounced peak during the Proterozoic(2 500 – 541 Ma), followed by a distinct fall during the Early Phanerozoic, and overall fluctuations through the Phanerozoic(541–0 Ma). Globally, seven stromatolite hotspots were identified from the new dataset, including western United States, eastern United States, western Europe, India, South Africa, northern China, and southern China. The proportional occurrences of inland aquatic stromatolites remain rather low(~20%) in comparison to marine stromatolites from the Precambrian to Jurassic, and then display a significant increase(30%–70%) from the Cretaceous to the present.
基金supported by the National Natural Science Foundation of China grants 61836014 to CL,and the STI2030‐Major Projects(2022ZD0205100)the Strategic Priority Research Program of Chinese Academy of Science,Grant No.XDB32010300+1 种基金Shanghai Municipal Science and Technology Major Project(Grant No.2018SHZDZX05)the Innovation Academy of Artificial Intelligence,Chinese Academy of Sciences to ZW.
文摘Lateral interaction in the biological brain is a key mechanism that underlies higher cognitive functions.Linear self‐organising map(SOM)introduces lateral interaction in a general form in which signals of any modality can be used.Some approaches directly incorporate SOM learning rules into neural networks,but incur complex operations and poor extendibility.The efficient way to implement lateral interaction in deep neural networks is not well established.The use of Laplacian Matrix‐based Smoothing(LS)regularisation is proposed for implementing lateral interaction in a concise form.The authors’derivation and experiments show that lateral interaction implemented by SOM model is a special case of LS‐regulated k‐means,and they both show the topology‐preserving capability.The authors also verify that LS‐regularisation can be used in conjunction with the end‐to‐end training paradigm in deep auto‐encoders.Additionally,the benefits of LS‐regularisation in relaxing the requirement of parameter initialisation in various models and improving the classification performance of prototype classifiers are evaluated.Furthermore,the topologically ordered structure introduced by LS‐regularisation in feature extractor can improve the generalisation performance on classification tasks.Overall,LS‐regularisation is an effective and efficient way to implement lateral interaction and can be easily extended to different models.
文摘Graph-based semi-supervised learning is an important semi-supervised learning paradigm. Although graphbased semi-supervised learning methods have been shown to be helpful in various situations, they may adversely affect performance when using unlabeled data. In this paper, we propose a new graph-based semi-supervised learning method based on instance selection in order to reduce the chances of performance degeneration. Our basic idea is that given a set of unlabeled instances, it is not the best approach to exploit all the unlabeled instances; instead, we should exploit the unlabeled instances that are highly likely to help improve the performance, while not taking into account the ones with high risk. We develop both transductive and inductive variants of our method. Experiments on a broad range of data sets show that the chances of performance degeneration of our proposed method are much smaller than those of many state-of-the-art graph-based semi-supervised learning methods.
基金supported by National key research and development program of China(2022YFA1503101)National Natural Science Foundation of China(22173067,22203058)+4 种基金Science and Technology Project of Jiangsu Province(BK20200873,BZ2020011)the Science and Technology Development Fund,Macao SAR(0052/2021/A)Collaborative Innovation Center of Suzhou Nano Science&Technology,the Priority Academic Program Development of Jiangsu Higher Education Institutions(PAPD)the 111 ProjectJoint International Research Laboratory of Carbon-Based Functional Materials and Devices。
文摘Lithium-sulfur(Li-S)batteries are notable for their high theoretical energy density,but the‘shuttle effect’and the limited conversion kinetics of Li-S species can downgrade their actual performance.An essential strategy is to design anchoring materials(AMs)to appropriately adsorb Li-S species.Herein,we propose a new three-procedure protocol,named InfoAd(Informative Adsorption)to evaluate the anchoring of Li_(2)S on two-dimensional(2D)materials and disclose the underlying importance of material features by combining high-throughput calculation workflow and machine learning(ML).In this paradigm,we calculate the anchoring of Li_(2)S on 12552D A_(x)B_(y)(B in the VIA/VIIA group)materials and pick out 44(un)reported nontoxic 2D binary A_(x)B_(y)AMs,in which the importance of the geometric features on the anchoring effect is revealed by ML for the first time.We develop a new Infograph model for crystals to accurately predict whether a material has a moderate binding with Li_(2)S and extend it to all 2D materials.Our InfoAd protocol elucidates the underlying structure-property relationship of Li_(2)S adsorption on 2D materials and provides a general research framework of adsorption-related materials for catalysis and energy/substance storage.
基金Project supptjrted by China Knowledge Centre for Engineering Sciences and Technology(CKCEST)。
文摘Recently,graph neural networks(GNNs)have achieved remarkable performance in representation learning on graph-structured data.However,as the number of network layers increases,GNNs based on the neighborhood aggregation strategy deteriorate due to the problem of oversmoothing,which is the major bottleneck for applying GNNs to real-world graphs.Many efforts have been made to improve the process of feature information aggregation from directly connected nodes,i.e.,breadth exploration.However,these models perform the best only in the case of three or fewer layers,and the performance drops rapidly for deep layers.To alleviate oversmoothing,we propose a nested graph attention network(NGAT),which can work in a semi-supervised manner.In addition to breadth exploration,a k-layer NGAT uses a layer-wise aggregation strategy guided by the attention mechanism to selectively leverage feature information from the k;-order neighborhood,i.e.,depth exploration.Even with a 10-layer or deeper architecture,NGAT can balance the need for preserving the locality(including root node features and the local structure)and aggregating the information from a large neighborhood.In a number of experiments on standard node classification tasks,NGAT outperforms other novel models and achieves state-of-the-art performance.
基金supported by the National Key Research and Development Program of China(2018YFB1600600)the National Natural Science Foundation of China under(61976034,U1808206)the Dalian Science and Technology Innovation Fund(2019J12GX035).
文摘The data generated from non-Euclidean domains and its graphical representation(with complex-relationship object interdependence)applications has observed an exponential growth.The sophistication of graph data has posed consequential obstacles to the existing machine learning algorithms.In this study,we have considered a revamped version of a semi-supervised learning algorithm for graph-structured data to address the issue of expanding deep learning approaches to represent the graph data.Additionally,the quantum information theory has been applied through Graph Neural Networks(GNNs)to generate Riemannian metrics in closed-form of several graph layers.In further,to pre-process the adjacency matrix of graphs,a new formulation is established to incorporate high order proximities.The proposed scheme has shown outstanding improvements to overcome the deficiencies in Graph Convolutional Network(GCN),particularly,the information loss and imprecise information representation with acceptable computational overhead.Moreover,the proposed Quantum Graph Convolutional Network(QGCN)has significantly strengthened the GCN on semi-supervised node classification tasks.In parallel,it expands the generalization process with a significant difference by making small random perturbationsG of the graph during the training process.The evaluation results are provided on three benchmark datasets,including Citeseer,Cora,and PubMed,that distinctly delineate the superiority of the proposed model in terms of computational accuracy against state-of-the-art GCN and three other methods based on the same algorithms in the existing literature.
文摘As an important method for knowledge graph(KG)complementation,link prediction has become a hot research topic in recent years.In this paper,a performance enhancement scheme for link prediction models based on the idea of semi-supervised learning and model soup is proposed,which effectively improves the model performance on several mainstream link prediction models with small changes to their architecture.This novel scheme consists of two main parts,one is predicting potential fact triples in the graph with semi-supervised learning strategies,the other is creatively combining semi-supervised learning and model soup to further improve the final model performance without adding significant computational overhead.Experiments validate the effectiveness of the scheme for a variety of link prediction models,especially on the dataset with dense relationships.In terms of CompGCN,the model with the best overall performance among the tested models improves its Hits@1 metric by 14.7%on the FB15K-237 dataset and 7.8%on the WN18RR dataset after using the enhancement scheme.Meanwhile,it is observed that the semi-supervised learning strategy in the augmentation scheme has a significant improvement for multi-class link prediction models,and the performance improvement brought by the introduction of the model soup is related to the specific tested models,as the performances of some models are improved while others remain largely unaffected.
基金supported by the NationalKeyR&DProgramof China underGrant No.2018YFA0701604.
文摘Due to the many types of distributed denial-of-service attacks(DDoS)attacks and the large amount of data generated,it becomes a chal-lenge to manage and apply the malicious behavior knowledge generated by DDoS attacks.We propose a malicious behavior knowledge base framework for DDoS attacks,which completes the construction and application of a multi-domain malicious behavior knowledge base.First,we collected mali-cious behavior traffic generated by five mainstream DDoS attacks.At the same time,we completed the knowledge collection mechanism through data pre-processing and dataset design.Then,we designed a malicious behavior category graph and malicious behavior structure graph for the characteristic information and spatial structure of DDoS attacks and completed the knowl-edge learning mechanism using a graph neural network model.To protect the data privacy of multiple multi-domain malicious behavior knowledge bases,we implement the knowledge-sharing mechanism based on federated learning.Finally,we store the constructed knowledge graphs,graph neural network model,and Federated model into the malicious behavior knowledge base to complete the knowledge management mechanism.The experimental results show that our proposed system architecture can effectively construct and apply the malicious behavior knowledge base,and the detection capability of multiple DDoS attacks occurring in the network reaches above 0.95,while there exists a certain anti-interference capability for data poisoning cases.
基金support of this research by the National Natural Science Foundation of China (No.52278117)the Philosophical and Social Science Program of Guangdong Province,China (GD22XGL20)the Shenzhen Science and Technology Program (No.20220531101800001 and No.20220810160221001).
文摘The continuous accumulation of operational data has provided an ideal platform to devise and implement customized data analytics for smart HVAC fault detection and diagnosis.In practice,the potentials of advanced supervised learning algorithms have not been fully realized due to the lack of sufficient labeled data.To tackle such data challenges,this study proposes a graph neural network-based approach to effectively utilizing both labeled and unlabeled operational data for optimum decision-makings.More specifically,a graph generation method is proposed to transform tabular building operational data into association graphs,based on which graph convolutions are performed to derive useful insights for fault classifications.Data experiments have been designed to evaluate the values of the methods proposed.Three datasets on HVAC air-side operations have been used to ensure the generalizability of results obtained.Different data scenarios,which vary in training data amounts and imbalance ratios,have been created to comprehensively quantify behavioral patterns of representative graph convolution networks and their architectures.The research results indicate that graph neural networks can effectively leverage associations among labeled and unlabeled data samples to achieve an increase of 2.86%–7.30%in fault classification accuracies,providing a novel and promising solution for smart building management.
