Search

Knowledge representation of scientific paper data is a problem to be solved,and how to learn the representation of paper nodes in scientific paper heterogeneous network is the core to solve this problem.This paper proposes an unsupervised cluster-level scientific paper heterogeneous graph node representation learning method(UCHL),aiming at obtaining the representation of nodes (authors,institutions,papers,etc.) in the heterogeneous graph of scientific papers.Based on the heterogeneous graph representation,this paper performs link prediction on the entire heterogeneous graph and obtains the relationship between the edges of the nodes,that is,the relationship between paper and paper.Experiments results show that the proposed method achieves excellent performance on multiple evaluation metrics on real scientific paper datasets.

With the development of science and technology,people need to access a large number of scientific and technological information quickly,and scientific paper is one of the main ways to carry scientific and technological information.As an important part of scientific paper,abstract is an effective tool for readers to retrieve literature.Therefore,the quality of abstract affects the retrieval rate of paper directly.However,due to the lack of writing experience,the quality of abstracts written by many authors is not high.Automatic generation of summary for scientific paper can help the author grasp the important content of paper more effectively,so as to write high-quality abstract.At the same time,the automatically generated abstract can also control the number of words in the abstract,which can bring more content to readers and help them understand the paper better.Generating automa-tic summarization for scientific paper can help author write abstract faster,which is one of the research contents in automatic summarization.Compared with common news document,scientific paper has the characteristics of strong structure and clear logical relationship.As far as the mainstream abstractive summarization such as encoder-decoder model is concerned,it mainly consi-ders the serialized information in the document,and rarely explores the text structure information in the document.For this reason,according to the characteristics in scientific papers,this paper proposes an automatic summarization model based on the hie-rarchical structure of "word-section-document",which uses the association between word and section to enhance the level of text structure and the interaction between levels,so as to screen out the key information in scientific paper.In addition,a context gate unit is extended to update the optimized context vector,thus capturing context information more comprehensively.The experimental results show that the proposed model can effectively improve the performance of the generated summarization in the ROUGE evaluation method.

With the rapid growth of urban populations,the number of private cars has grown exponentially,which makes overwhelming traffic congestion problem become more and more acute.The traditional traffic signal control technology is difficult to adapt to the complex and changeable traffic conditions,and the data-driven methods bring new research directions for the control-based system.The combination of deep reinforcement learning and traffic control systems plays an important role in adaptive traffic signal control.First,this paper reviews the latest progress in the application of intelligent traffic signal control systems,the methods of intelligent traffic signal control are classified and discussed,and the existing works in this field are summarized.The deep reinforcement learning method can effectively solve the problems of inaccurate state information acquisition,poor algorithm robust and weak regional coordination control ability in intelligent traffic signal control.Then,on the basis of the above,this paper gives an overview of the simulation platforms and experimental setup for intelligent traffic signal control,and analyzes and verifies it through examples.Finally,The challenges and unsolved problems in this field are discussed and future research directions are summarized.

Biomarker recognition is a major basis for achieving precision medicine,which plays an important role in diagnosing complex diseases,judging disease stages and evaluating the safety and effectiveness of new drugs or therapies in the target population.As the key technology of intelligent identification of biomarkers,the prediction of the association between circular RNA and disease is the key to deeply evaluate and measure the biological process,pathological process and intervention pathological response of subjects,and is one of the effective means and approaches to practice “precision medicine”.This paper comprehensively combs and prospects the circular RNA disease association prediction model in biological big data.Specifically,it first discusses the relationship between circRNA and disease in terms of the research background,physicochemical properties and functions of circ-RNAs.Then,it investigates the public database resources of circRNAs and diseases,and summarizes four computational methods of circRNA-disease association prediction from the perspective of computational models,and analyzes their advantages and shortcomings.Finally,this paper discusses the current challenges and future possible research directions of circRNA-disease association prediction problem.

As a new network architecture,the software defined network(SDN)simplifies the network management logic by separating data plane and control plane,which is one of the popular research subjects of next-generation network.However,due to frequent link failures and other factors,it is difficult to guarantee the reliability of SDN,which is a problem well recognized in the industry.The existing SDN link failure recovery models often have the problems of long recovery delay,requiring too many flow entries and ignoring link congestion after recovery from failure.To solve these problems,this paper proposes a single-link failure recovery model(LFA-CA)based on segment routing(SR).The model employs the two heuristic algorithms of BPF and BPU to calculate a loop-free backup path during network initialization and update the congestion avoiding backup path during operation,respectively,so as to achieve fast recovery from single-link failure and congestion avoidance after failure.In this paper,massive simulation experiments are carried out to evaluate the performance of our model,and the results prove that compared with some of the existing SDN single-link failure recovery models,LFA-CA consumes less forwarding rules and has better load balancing ability after failure.

Since 2018,scholars have proposed and studied a new topic of “concept reduction” in formal concept analysis.Including unnecessary concepts,core concepts,relatively necessary concepts,and the identification of three types of concepts,and research on concept reduction algorithm.In this paper,the same effect relation is proposed and its important properties are studied.It pre-sents a simple way to identify three types of concepts through the same effect relationship,and proposes a new algorithm for concept reduction which is based on the concept lattice of the complement set of subsets of the same effect relationship.For decades,the algorithm of “reduction topic” has adopted the method of conjunctive normal form and disjunctive normal form transformation.Many scholars even said that “reduction problem” was equivalent to the transformation of conjunctive paradigm and disjunctive paradigm.This paper studies a new method to solve the “reduction problem” without using the transformation between conjunctive normal form and disjunctive normal form.This new method is of significance both in theory and in practice.It is a new attempt.There are often many “concept reduction” in a background,so it is not very meaningful to find out all the results.Gene-rally,it is necessary to find out “concept reduction” containing some concepts,and the method in this paper has particular advantages in this respect.

In the past decade,widespread data collection has become the norm.With the rapid development of large-scale data analysis and machine learning,data privacy is facing fundamental challenges.Exploring the trade-offs between privacy protection and data collection and analysis is a key scientific question.Differential privacy has become a de facto data privacy standard and has been widely studied and applied.Differential privacy technology can provide strict privacy protection for user data through a certain randomization mechanism.This paper provides a comprehensive overview of differential privacy technology and a summary and analysis of the latest progress of differential privacy.Specifically,this paper first gives a theoretical summary of differential privacy,including the central model,the local model and the shuffle model proposed in recent years.The three models are compared,and the advantages and disadvantages of different models are analyzed.Then,on the basis of the three models,some typical differential privacy mechanisms in literatures are analyzed from the perspective of algorithms.Then the current application of differential privacy technology in various fields is introduced.Finally,some new research topics about differential privacy technology are introduced,which expand the rich research direction for differential privacy technology.

With the rapid development and integration of biomedicine and information technology,massive amounts of imaging data,patient report data,electronic health records,and omics data have been accumulated rapidly in healthcare.These data are cha-racterized by complexity,heterogeneity and high dimensionality.Deep learning has the ability of complex function simulation and automatic feature learning,which can provide efficient technical support for research in medical diagnosis and drug development.Currently,deep learning has been extremely successful in medical imaging and further more,some medical imaging diagnostic systems based on deep learning have achieved performance that is even comparable to that of relevant experts.Due to the progress of natural language processing technology,deep learning has also made remarkable progress in the use of non-image data tasks.This paper first briefly describes the development of deep learning in healthcare.Subsequently,the application of deep learning model in healthcare is statistically analyzed,and some available datasets are sorted out.In addition,this paper also introduces the research progress of deep learning in medical diagnosis and treatment processes such as disease diagnosis and health monitoring,and its research progress in protein structure prediction and drug discovery.Finally,key challenges of deep learning in healthcare applications such as data quality,interpretability,privacy security and practical application limitations are discussed.It also discusses feasible solutions or approaches to these challenges.

Batched matrix computing problems are widely existed in scientific computing and engineering applications.With rapid performance improvements,GPU has become an important tool to solve such problems.The eigenvalue decomposition belongs to the two-sided decomposition and must be solved by the iterative algorithm.Iterative numbers for different matrices can be varied.Therefore,designing eigenvalue decomposition algorithms for batched matrices on the GPU is more challenging than designing batched algorithms for the one-sided decomposition,such as LU decomposition.This paper proposes batched algorithms based on the Jacobi algorithms for eigenvalue decomposition of Hermitian matrices.For matrices that cannot reside in shared memory wholly,the block technique is used to improve the arithmetic intensity,thus improving the use of GPU resources.Algorithms presented in this paper run completely on the GPU,avoiding the communication between the CPU and GPU.Kernel fusion is adopted to decrease the overhead of launching kernel and global memory access.Experimental results on V100 GPU show that our algorithms are better than existing works.Performance evaluation results of the Roofline model indicate that our implementations are close to the upper bound,approaching 4.11TFLOPS.

Knowledge Graph-to-Text is a new task in the field of knowledge graph,which aims to transform knowledge graph into readable text describing these knowledge.With the deepening of research in recent years,the generation technology of Graph-to-Text has been applied to the fields of product review generation,recommendation explanation generation,paper abstract generation and so on.The translation model in the existing methods adopts the method of first-plan-then-realization,which fails to dynamically adjust the planning according to the generated text and does not track the static content planning,resulting in incohe-rent semantics before and after the text.In order to improve the semantic coherence of generated text,a Graph-to-Text model based on dynamic memory and two-layer reconstruction enhancement is proposed in this paper.Through three stages of static content planning,dynamic content planning and two-layer reconstruction mechanism,this model makes up for the structural difference between knowledge graph and text,focusing on the content of each triple while generating text.Compared with exis-ting generation models,this model not only compensates for the structural differences between knowledge graphs and texts,but also improves the ability to locate key entities,resulting in stronger factual consistency and semantics in the generated texts.In this paper,experiments are conducted on the WebNLG dataset.The results show that,compared with the current exis-ting models in the task of Graph-to-Text,the proposed model generates more accurate content planning.The logic between the sentences of the generated text is more reasonable and the correlation is stronger.The proposed model outperforms existing methods on me-trics such as BLEU,METEOR,ROUGE,CHRF++,etc.

Cross-item social recommendation is a method of integrating social relationships into the recommendation system.In social recommendation,user is the bridge connecting user-item interaction graph and user-user social graph.So user representation learning is essential to improve the performance of social recommendation.However,existing methods mainly use static attributes of users or items and explicit friend information in social networks for representations learning,and the temporal information of the interaction between users and items and their implicit friend information are not fully utilized.Therefore,in social recommendation,effective use of temporal information and social information has become one of the important research topics.This paper focuses on the temporal information of the interaction between users and items,and gives full play to the advantages of social network,modeling the user's implicit friends and item's social attributes.This paper proposes a novel graph neural networks social recommendation based on high-order and temporal features,referred to as HTGSR.Firstly,the framework uses gated recurrent unit to model item-based user representations to reflect the user's recent preferences,and defines a high-order mo-deling unit to extract the user's high-order connected features and obtain the user's implicit friend information.Secondly,HTGSR uses attention mechanism to obtain social-based user representation.Thirdly,the paper proposes different ways to construct item's social networks,and uses the attention mechanism to obtain item representations.Finally,the user's and item's representations are input to the MLP to complete the user's rating prediction for the item.The paper conducts specific experiments on two public and real-world datasets,and compares the experimental results with different recommendation algorithms.The results show that the HTGSR has achieved good results on the two datasets.

Real-time collision avoidance in dynamic environments is a challenge in trajectory planning of mobile robots. Focusing on environments with variable number of obstacles,this paper proposes a real-time trajectory planning algorithm,Crit-LSTM-DRL,based on long short-term memory(LSTM) and deep reinforcement learning(DRL). First,it predicts the time to the occurrence of a collision between an obstacle and the robot based on their states,and then computes the collision criticality of each obstacle with respect to the robot. Second,it generates the obstacle sequence based on the collision criticality and abstracts a fixed-dimension vector by LSTM to represent the environment. Finally,the robot state and the extracted vector are concatenated as the input of the DRL's value network to compute the value with respect to the system state. At any instant,for each action,it predicts the value of the next state based on the LSTM and DRL models and then the value of the current state; hence,the action generating the maximal value of the current state is selected to control the robot. To evaluate the performance of Crit-LSTM-DRL,it is first trained in three different environments and obtain three models: the model trained in the environment with 5 obstacles,the model trained in the environment with 10 obstacles,and the model trained in the environment with variable number of obstacles(1～10). The models then are tested in various environments containing different number of obstacles. To further investigate the effects of the interaction between an obstacle and the robot,this paper also takes the joint state of an obstacle and the robot as the state of the obstacle and trains another three models in the above training environments. Experimental results show the effectiveness and efficiency of Crit-LSTM-DRL.

The existing oil painting art style classification algorithms ignore the influence of the main area and the overall effect on the art style.Aiming at this problem,this paper proposes a new oil painting classification algorithm based on multi-feature fusion classifier(MFFC).Firstly,based on the common arrangement form of oil painting art elements,this paper designs the overlapping image block method.This method extracts spatial features of oil paintings to make up for the lack of composition style in existing algorithms.And it also can be used to distinguish the subject area from the background area.Secondly,the spatial features and the underlying features are combined in series to increase the location information of the elements in the picture.Finally,the spatial voting method is designed to give priority to the classification result of the main area as the output result of the algorithm.This is to highlight the role of oil painting subject area in the classification and realize the automatic classification of oil painting art style.Tested on the data set created by the FS-classifier model,its accuracy,precision,recall,F1-score and AUC reaches 96.92%,63.69%,98.75%,98.57% and 0.917,respectively.Compared with FS-classifier,the result increases by 6.72%,5.85%,9.05%,7.1% and 0.128,respectively.When tested on WIKIART and compared with other six algorithms,the accuracy improves by 13.27%,at least.The results show that the proposed algorithm can effectively improve the performance of spatial features for oil painting art style classification task,and has good practical value.

In recent years,artificial intelligence represented by deep learning has made breakthroughs in theories and technologies.With the strong support of data,algorithms and computing power,deep learning has received unprecedented attention and has been widely used in various fields,bringing great improvements to the corresponding fields.With the wide application of deep learning technology in various fields including security critical ones,the security issue of deep learning has attracted more and more attention.Researchers have found many security risks in deep learning systems.In terms of the security of deep learning models,researchers have extensively explored the new attack paradigm of backdoor attack.Backdoor attack can threaten deep learning models throughout their whole life cycle.A large number of researchers have proposed series of attack scheme from different angles.This paper takes the security threats of deep learning system as a starting point,introduces the current attack paradigms.On this basis,it gives the back-ground and principle of backdoor attack,distinguishes the similar attack paradigms such as adversarial attack and data poisoning attack,then continues to elaborate on the attack principle and outstanding features of the classic methods of backdoor attack to date.According to the working principle,the attack schemes are divided into data poisoning based attack and model poisoning based attack and others,the paper systematically summarizes them and clarify the advantages and disadvantages of current research.Then,this paper surveys the state-of-the-art works of backdoor attack against various typical applications and popular deep learning paradigms,which further reveal the threat of backdoor attack towards deep learning models.Finally,this paper summarizes the research work on applying backdoor attack characteristics to positive applications and explores the current challenges of backdoor attack,as well as discusses future research directions worthy of in-depth exploration,aiming to provide guidance for the follow-up researchers to further promote the development of backdoor attack and security of deep learning.

Compared with English text which is naturally composed of words,Chinese text has no word delimiters,so the combination of Chinese characters is more flexible,and it's more difficult to determine the entity boundaries in Chinese named entity recognition(NER).Current mainstream methods transform the NER task into a sequence labeling task.This paper studies the predicted label sequence under the BIOES tag scheme and calculates the entity boundary accuracy by separately considering the entity head label B or tail label E,which shows that increasing the boundary accuracy can further improve the accuracy of entity recognition.We expand the boundaries of entities with continuous labels,use the label type of the last character of the entity to correct the entity type,and use the word segmentation information to fill in the entity with incomplete labels.Finally,this paper proposes a BIO+ES labeling scheme that adds boundary labels to distinguish non-entity characters at entity boundaries and further improves the performance of Chinese NER.

Knowledge graphs describe objective knowledge in the real world in a structured form,and are confronted with issues of completeness and newly-added knowledge.As an important means of complementing and updating knowledge graphs,know-ledge graph reasoning aims to infer new knowledge based on existing knowledge.In recent years,the research on knowledge graph reasoning based on repre-sentation learning has received extensive attention.The main idea is to convert the traditional reasoning process into semantic vector calculation based on the distributed representation of entities and relations.It has the advantages of fast calculation efficiency and high reasoning performance.In this paper,we review the knowledge graph reasoning based on repre sentation learning.Firstly,this paper summarizes the symbolic representation,data set,evaluation metric,training method,and evaluation task of knowledge graph reasoning.Secondly,it introduces the typical methods of knowledge graph reasoning,including translational distance and semantic matching methods.Thirdly,multi-source information fusion-based knowledge graph reasoning methods are classified.Then,neural network-based reasoning methods are introduced including convolutional neural network,graph neural network,recurrent neural network,and capsule network.Finally,this paper summarizes and forecasts the future research direction of knowledge graph reasoning.

The ranking of nodes’ influence has always been a hot issue in the research area of complex networks.Susceptible-infected-recovered(SIR) model is an ideal nodes’ influence ranking method,which is commonly used to evaluate other nodes’ in-fluence ranking methods.But it is difficult to be applied in practice due to its high time complexity.This paper proposes a nodes’ influence ranking model based on sir value learning.Both the local structure and global structure information of nodes are used as features in the model.The sir value learning model is constructed by means of a deep learning model,which is trained on nodes’ features and sir data set in synthetic graphs with the same size.The trained model can predict sir value based on nodes’ features,and then rank nodes’ influence based on predicted sir.In this paper,a specific nodes’ influence ranking method is implemented based on the proposed model,and experiments are carried out on five real networks to verify the effectiveness of the method.The results show that the accuracy and monotonicity of nodes’ influence ranking results are improved compared with degree centrality,Kshell and Weighted Kshell degree neighborhood.

Traditional IP-based Internet offers an end-to-end data transport service and has developed rapidly in the past half-century.However,serious security incidents emerged from attacks based on traditional networks.Traditional security mechanisms(e.g.,firewalls,intrusion detection systems) enhance security.However,most of them only provide some remedial strategies rather than solve the address-security problem radically due to the lack of change in network design.The overall in-depth security of the networked system cannot be guaranteed without a fundamental change.In order to meet the development requirements of the next generation of an endogenous security network,one of the future networks,the multi-identifier network(MIN),is introduced as our research background.This paper proposes an efficient scheme in hieratical architecture that provides comprehensive protection by addressing the security aspects pertaining to the network and application layers.At the network layer,the proposed architecture develops a multi-identifier routing scheme with embedded identity-based authentication and packet signature mechanisms to provide data tamper-resistance and traceability.At the application layer,the proposed architecture designs a mimic defensive scheme combined with weighted network centrality measures.This scheme focuses on protecting the core components of the whole network to improve the service's robustness and efficiently resist potential attacks.This paper tests and evaluates the proposed scheme from a theoretical and practical perspective.An analytical model is built based on the random walk for theoretical evaluation.In experiments,the proposed scheme is developed in MIN as MIN-VPN.Then considering IP-VPN as a baseline,anti-attack tests are conducted on IP-VPN and MIN-VPN.The results of theoretical evaluations and experiments show that the proposed scheme provides excellent transmission performance and successful defense against various TCP/IP-based attacks with acceptable defensive cost,demonstrating this security mechanism's effectiveness.In addition,after long-period penetration testing in three international elite security contests,the proposed method is effectively immune to all TCP/IP-based attacks from thousands of professional teams,thus verifying its strong security.

In recent years,machine learning has not only achieved remarkable results in conventional fields such as computer vision and natural language processing,but also been widely applied to process sensitive data such as face images,financial data and medical information.Recently,researchers find that machine learning models will remember the data in their training sets,making them vulnerable to membership inference attacks,that is,the attacker can infer whether the given data exists in the training set of a specific machine learning model.The success of membership inference attacks may lead to serious individual privacy leakage.For example,the existence of a patient's medical record in a hospital's analytical training set reveals that the patient was once a patient there.The paper first introduces the basic principle of membership inference attacks,and then systematically summarizes and classifies the representative research achievements on membership inference attacks and defenses in recent years.In particular,how to attack and defend under different conditions is described in detail.Finally,by reviewing the development of membership inference attacks,this paper explores the main challenges and potential development directions of machine learning privacy protection in the future.

Cracks in the concrete pavement will affect the safety,applicability,and durability of the structure,and crack detection is a challenging research hotspot.This paper proposes a crack detection model composed of an improved full convolutional network and a deep supervision network,which uses the improved VGG-16 as the backbone network.Firstly,the low-level convolutional feature aggregation is fused to the backbone network again through the spatial attention mechanism.Secondly,the middle and high-level convolutional features are fused through the lightweight dilated convolution fusion module for multi-feature fusion to get the clear edge and high-resolution feature maps,all side feature maps are added to produce the final prediction map.Finally,the deep supervision network provides direct supervision for the detection results of each stage.In this paper,the focus loss function is selected as the evaluation function,and the trained network model can efficiently identify the crack location from the input original image under various conditions such as uneven illumination and complex background.To verify the effectiveness and robustness of the proposed method,it is compared with six methods on three datasets,DeepCrack,CFD,and Crack500,and the results show that it has excellent performance,and the F-score value reaches 87.12%.

Based on the idea of crowdsourcing,mobile crowdsensing(MCS) collects mobile sensing devices to sense the surroun-ding environment,which can make environment sensing and information collection more flexible,convenient and efficient.Whe-ther the task allocation strategy is reasonable or not directly affects the success of the sensing task.Therefore,formulating a reasonable task allocation strategy is a hotspot and focus in the research of MCS.At present,most of the task allocation methods in MCS systems are offline and targeted at single type tasks.However,in practice,online multi-type task allocation is more common.Therefore,this paper studies the task allocation method in MCS for multiple types of tasks,and proposes an online task allocation strategy oriented to system benefits combined with the characteristics of MCS technology in the military field.In this paper,a long-term,dynamic online task allocation system model is established,and the problem is solved based on Lyapunov optimization theory with the system benefit as the optimization goal,so that the online dynamic control of task admission strategy and task allocation scheme is realized.Experiment shows that the online task allocation algorithm proposed in this paper is effective and feasible.It can reasonably allocate the tasks arriving at the MCS system online,ensure the stability of the task queue,and increase the system utility by adjusting the parameter value.

In a multi-robot environment,robots with different capabilities collaborate with each other to complete task requirements.Realistically,these tasks are dynamically issued and can have different goals and urgency levels,so it is necessary to allocate and schedule the appropriate robots responsible for executing the fine-grained actions decomposed for each task.Most of the existing approaches are suitable for static and homogeneous task allocation scenarios,while most of the dynamic heterogeneous tasks are assigned using exclusive allocation strategies,which causes the robot to frequently enter into waiting states(i.e.,robots are in the idle phase between being assigned a task and actually starting to execute it).Since tasks vary in urgency levels and release times,this allocation method will reduce the efficiency of response to more urgent tasks,while leading to longer waiting time and task completion time.To address this problem,this paper proposes a fine-grained action allocation and scheduling method for dynamic heterogeneous tasks in a multi-robot environment.In this paper,the object of allocation and scheduling is a fine-grained action decomposed by a task,and an action can be undertaken by one capability of a robot.Faced with a set of fine-grained actions decomposed by the task,this method draws on the auction algorithm process to calculate the optimal allocation scheme for a robot to undertake a specific action based on the robot capabilities,state and task information.In addition,by executing the allocation and scheduling process at each new task release or when a robot finishes executing an action,robots in the general task waiting state can be scheduled to the urgent task to ensure the priority completion of the urgent task and reduce the overall waiting time of the robot.Based on this approach,the execution module ROSPlan is extended and implemented.Simulation experiments around a set of multi-robot dynamic heterogeneous tasks show that the proposed methodr can obtain a better allocation scheme compared to the method using greedy policies.

As a new biometric technology,gait recognition has become a research hotspot of biometric technology due to its advantages of non-contact,non-interference,long-distance and difficult to camouflage.In addition,in recent years,due to the maturity of MEMS inertial sensor technology and its wide application in portable devices,gait recognition based on inertial sensors has attracted more and more attention from researchers.This paper collects and sorts out the research methods and research status of gait recognition using inertial sensors at home and abroad,and reviews the relevant technologies in this field.According to the sequence of recognition processing,the paper reviews the relevant technologies and research status of data acquisition,data preprocessing,data segmentation,feature selection and combination,and intelligent recognition in each stage.At the same time,the main public gait databases are given for the convenience of interested readers.Finally,on this basis,the technical difficulties of gait-based recognition are discussed,and the future development directions are prospected.

Bayesian inference is one of the main problems in statistics.It aims to update the prior knowledge of the probability distribution model based on the observation data.For the posterior probability that cannot be observed or is difficult to directly calculate,which is often encountered in real situations,Bayesian inference can obtain a good approximation.It is a kind of important method based on Bayesian theorem.Many machine learning problems involve the process of simulating and approximating the target distribution of various types of feature data,such as classification models,topic modeling,and data mining.Therefore,Bayesian inference has shown important and unique research value in the field of machine learning.With the beginning of the big data era,the experimental data collected by researchers through actual information is very large,resulting in the complex distribution of targets to be simulated and calculated.How to perform accurate and time-efficient approximation inferences on target distributions under complex data has become a major and difficult point in Bayesian inference problems today.Aiming at the infe-rence problem under this complex distribution model,this paper systematically introduces and summarizes the two main methods for solving Bayesian inference problems in recent years,which are variational inference and sampling methods.Firsly,this paper gives the problem definition and theoretical knowledge of variational inference,introduces in detail the variational inference algorithm based on coordinate ascent,and gives the existing applications and future prospects of this method.Next,it reviews the research results of existing sampling methods at home and abroad,gives the specific algorithm procedure of various main sampling methods,as well as summarizes and compares the characteristics,advantages and disadvantages of these methods.Finally,this paper introduces parallel tempering technique,outlines its basic theories and methods,discusses the combination and application of parallel tempering and sampling methods,and explores new research directions for the future development of Bayesian inference problems.

Recently,the fairness of the algorithm has aroused extensive discussion in the machine learning community.Given the widespread popularity of spectral clustering in modern data science,studying the algorithm fairness of spectral clustering is a crucial topic.Existing fair spectral clustering algorithms have two shortcomings:1) poor fairness performance;2) work only for single sensitive attribute.In this paper,the fair spectral clustering problem is regarded as a constrained spectral clustering problem.By solving the feasible solution set of constrained spectral clustering,an unnormalized fair spectral clustering(UFSC) method is proposed to improve fairness performance.In addition,the paper also proposes a fair clustering algorithm suitable for multiple sensitive attribute constraints.Experimental results on multiple real-world datasets demonstrate that the UFSC and MFSC are fairer than the existing fair spectral clustering algorithms.

As a new architecture,the space-air-ground integrated communication technology can effectively improve the network service quality of ground terminal,and has attracted widespread attention in recent years.This paper studies a space-air-ground integrated mobile edge computing system,in which multiple UAVs provide low-latency edge computing services for ground devices,and low earth orbit satellites provide ubiquitous cloud computing services for ground devices.Since the deployment position of the UAVs and the scheduling scheme of computing tasks are the key factors affecting the performance of the system,the deployment position of the UAVs,the link relationship between the ground terminal and the UAVs,and the offloading ratio of computing tasks need to be jointly optimized to minimize the average task response delay of the system.Since the formally defined joint optimization problem is a mixed nonlinear programming problem,this paper designs a two-layer optimization algorithm.In the upper layer of the algorithm,a particle swarm optimization algorithm that combines genetic algorithm operators is proposed to optimize the deployment position of the UAVs,and the greedy algorithm is used in the lower layer of the algorithm to optimize the computing task offloading scheme.The extensive simulation experiments verify the feasibility and effectiveness of the proposed method.The results show that the proposed method can achieve lower average task response time,compared to other baseline methods.

This paper mainly focuses on the mean estimation problem in differential privacy query.After introducing the current mainstream local differential privacy design scheme of numerical data mean estimation,it first introduces the random censoring mechanism in random response technology to reveal the basic principle of mean calculation under local differential privacy,proposes a utility optimization theorem about the variance of mean estimation,and gives a boundary optimization formula,which improves the interpretability and operability of utility optimization theory in this field.Based on this theory,this paper proposes a practical,concise and efficient mean estimation algorithm protocol RCP for the first time,which can be used to collect and analyze the data of intelligent device users connected to the Internet,while meeting the requirements of local differential privacy.RCP is simple in structure,supports data analysis tasks on any number of numerical attributes,and has efficient communication and calculation,effectively alleviating the practical problems of complex algorithm design,difficult optimization,and low efficiency.Finally,empirical research demonstrates that the proposed method outperforms other existing schemes in terms of utility,efficiency and asymptotic error bounds.

Artificial intelligence has been integrated into all aspects of people's daily lives with the continuous development of machine learning,especially in the deep learning area.Machine learning models are deployed in various applications,enhancing the intelligence of traditional applications.However,in recent years,research has pointed out that personal data used to train machine learning models is vulnerable to the risk of privacy disclosure.Membership inference attacks(MIAs) are significant attacks against the machine learning model that threatens users' privacy.MIA aims to judge whether user data samples are used to train the target model.When the data is closely related to the individual,such as in medical,financial,and other fields,it directly interferes with the user's private information.This paper first introduces the background knowledge of membership inference attacks.Then,we classify the existing MIAs according to whether the attacker has a shadow model.We also summarize the threats of MIAs in different fields.Also,this paper points out the defense means against MIAs.The existing defense mechanisms are classified and summarized according to the strategies for preventing model overfitting,model-based compression,and disturbance.Finally,this paper analyzes the advantages and disadvantages of the current MIAs and defense mechanisms and proposes possible research directions for future MIAs.

The task of object detection is a basic and highly concerned work in the field of computer vision.Because object detection in remote sensing has important application value in transportation,military,agriculture,etc.,it has also become a major research hotspot.Compared with natural images,remote sensing images are affected by many factors such as complex background interference,weather,irregularities,and small objects.It is extremely challenging to achieve higher accuracy in remote sensing image object detection tasks.This paper proposes a novel object detection network based on Transformer,swin pyramid Transformer(SPT).SPT uses a sliding window Transformer module as the backbone of feature extraction.Among it,the self-attention mechanism of Transformer is very effective for detecting objects in a chaotic background,and the sliding window mode efficiently avoids a large number of square-level complexity calculations.After obtaining the feature map extracted by the backbone network,SPT uses a pyramid architecture to fuse different scale and semantic features,pithily reducing the loss of information between feature layers,and capturing the inherent multi-scale hierarchical relationship.In addition,this paper proposes self-mixed Transformer(SMT) module and cross-layer Transformer(CLT) module.SMT re-renders the highest-level feature map to enhance object feature recognition and expression.According to the feature context interaction,the feature expressions of the pixels of each feature layer are rearranged by CLT,and the CLT module is integrated into the bottom-up and top-down dual paths of the pyramid to make full use of global and local information containing different semantics.Our SPT network model is trained and tested on the UCAS-AOD and RSOD datasets.Experimental results show that SPT is high-performing in remote sensing image object detection tasks,especially suitable for irregular and small target categories,such as overpass and car.

With the increasingly prominent limitations of deep neural networks in practical applications,brain-like computing spiking neural networks with biological interpretability have become the focus of research.The uncertainty and complex diversity of application scenarios pose new challenges to researchers,requiring brain-like computing spiking neural networks with multi-scale architectures similar to biological brain organizations to realize the perception and decision-making function of multi-modal and uncertain information.This paper mainly introduces the multi-scale biological rational brain-like computing spiking neural network model and its learning algorithm for multi-modal information representation and uncertainty information perception,analyzing and discussing two key technical issues that the spiking neural network based on the interconnection of memristors can rea-lize multi-scale architecture brain-like computing,namely:the consistency problem of multi-modal and uncertain information with spike timing representation,and the computing fault-tolerant problem for the multi-scale spiking neural network with different learning algorithms.Finally,this paper analyzes and forecasts the further research direction of brain-like computing spiking neural network.

As an important branch in the field of computer vision,object tracking has been widely used in many fields such as intelligent video surveillance,human-computer interaction and autonomous driving.Although object tracking has achieved good development in recent years,tracking in complex environment is still a challenge.Due to problems such as occlusion,object deformation and illumination change,tracking performance will be inaccurate and unstable.In this paper,an effective object tracking method AFTM,is proposed with attention features.Firstly,this paper constructs an adaptively generated attention weight factor group,which implements an efficient adaptive fusion strategy for response map to improve the accuracy of object positioning and bounding box scale calculation in the process of classification and regression.Secondly,aiming at the class imbalance in the data set,the proposed method uses the dynamically scaled cross entropy loss as the loss function of the object positioning network,which can modify the optimization direction of the model and make the tracking performance more stable and reliable.Finally,this paper designs a corresponding learning rate adjustment strategy to stochastically average the weight of a number of models,which can enhance the generalization ability of the model.Experimental results on public data sets show that the proposed method has higher accuracy and more stable tracking performance in complex tracking environment.

Clustering analysis plays a key role in machine learning,data mining and biological DNA information.Clustering algorithms can be categorized into flat clustering and hierarchical clustering.Flat clustering mostly divides the data set into K parallel communities without intersections,but the real communities have multi-level inclusion relations,so the hierarchical clustering algorithms can provide more elaborate analysis and better interpretability.Compared with flat clustering,the research progress of hierarchical clustering is slow.Aiming at the problem of hierarchical clustering,this paper surveys a large number of related papers in the aspects of the selection of cost functions,the evaluations of clustering results and the performance of clustering algorithms.The evaluation indices of clustering results mainly include modularity,Jaccard index,normalized mutual information,dendrogram purity,etc.Among the flat clustering algorithms,the classical algorithms include K-means algorithm,label propagation algorithm,DBSCAN algorithm,spectral algorithm and so on.The hierarchical clustering algorithms can be further classified into agglomerative clustering algorithms and split clustering algorithm.The split clustering algorithms involves dichotomy K-means algorithm,recursive sparsest cut algorithm,etc.Agglomerative clustering algorithms involves classical Louvain algorithm,BIRCH algorithm and more recent HLP algorithm,PERCH algorithm and GRINCH algorithm.This paper also further analyzes the advantages and disadvantages of these algorithms,and finally,the whole paper is summarized.

Computation graph simplification is a kind of optimization technique to improve the training speed of graph neural network models.It uses the characteristics of common neighbors among nodes and speeds up the training of graph neural network models by eliminating redundant computation in the stage of aggregation.However,when dealing with large-scale graph data,the existing computation graph simplification techniques suffer from the problem of low computation efficiency,which affects the application of computation graph simplification in large-scale graph neural networks.This paper analyzes the current techniques of computation graph simplification in detail by counting the overhead of two phases including searching and reconstruction,and summarizes the shortcomings of existing techniques.On this basis,it proposes an algorithm of fast computation graph simplification via influence-based pruning for graph neural network.This algorithm applies an influence model to describe the contribution of each node to the computation graph simplification and prunes the searching space of common neighbors based on influence,which greatly improves the efficiency of the phase of searching.In addition,this paper analyzes the algorithm complexity and theoretically proves the expected acceleration effect of this technique.Finally,in order to verify the effectiveness of this novel algorithm,the algorithm is applied to two mainstream computation graph simplification technique,and common graph neural network models areselected to test on some data sets.Experimental results demonstrate that the novel algorithm can significantly improve the efficiency of the computation graph simplification on the premise of ensuring a certain amount of redundant computation reduction.Compared with the baseline of computation graph simplification,the proposed technique can speed up to 3.4 times in searching phase and speed up to 1.6 times on the whole process on PPI dataset,while it can speed up to 5.1 times in searching phase and speed up to 3.2 times on the whole process on Reddit dataset.

With the development of computer technology,using machine learning algorithms to build automated evaluation models has become an important tool to for the financial institutions to conduct credit evaluation.However,currently,the credit evaluation model is still facing challenges:credit data is class-imbalanced and high-dimensional,meanwhile,the behavior of customers can be influenced by the changeable external environment,namely,the concept drift will occur.As a result,this paper proposes a dynamic credit evaluation model,which can achieve the flexible model update by using ensemble learning algorithm to continuously add base classifiers trained on new incremental data,and dynamically adjusting the weight of each base classifier to adapt to concept drift.When concept drift occurs,according to the detection results of concept drift,the model is able to use different forms of equalization and feature selection on credit data.In particular,for feature selection,this paper proposes an incremental feature selection algorithm combining the choice of representative samples that makes the feature selection efficient and accurate,enabling the model to simultaneously process the high-dimensional imbalanced data and adapt the concept drift of incremental credit data.Finally,this paper manages to demonstrate that the proposed dynamic model is more efficient and accurate than other prevailing algorithms on real incremental high-dimensional credit datasets.

Blockchain is a distributed database based on the Bitcoin platform.It has attracted widespread attention from academia and industry because of its characteristics of decentralization,traceability,and non-tamperability.However,with the rapid growth of the number of nodes and the amount of data,problems such as low throughput and difficult storage scalability appear.The problem of storage scalability of blockchain has become a bottleneck for blockchain applications.Therefore,this paper focuses on the summary and prospect of storage scalability of blockchain.Firstly,this paper introduces the basic knowledge and data structure of blockchain and analyzes the problems of storage scalability of blockchain at present.Secondly,this paper explains the principles,implementation methods,research progress,advantages,and disadvantages of the 0th layer expansion scheme,the on-chain expansion scheme,and the off-chain expansion scheme.Finally,this paper analyzes the challenges in the current blockchain sto-ragescalability scheme and provides a direction for the next blockchain research work.

GCN-based collaborative filtering models generate the representation of user nodes and item nodes by aggregating information on user-item interaction bipartite graph,and then predict users' preferences on items.However,they neglect users' different interaction intents and cannot fully explore the relationship between users and items.Existing graph disentangled collaborative filtering models model users' interaction intents,but ignore the global information of interaction graph and the essential features of users and items,causing the incompleteness of representation semantics.Furthermore,disentangled representation learning is inefficient due to the iterative structure of model.To solve these problems,this paper devises a deep disentangled collaborative filtering model incorporating graph global information,which is named as global graph disentangled collaborative filtering(G2DCF).G2DCF builds graph global channel and graph disentangled channel,which learns essential features and intent features,respectively.Meanwhile,by introducing orthogonality constraint and representation independence constraint,G2DCF makes every user-item interaction intent as unique as possible to prevent intent degradation,and raises the independence of representations under different intents,so as to improve the disentanglement effect.Compared with the previous graph collaborative filtering models,G2DCF can more comprehensively describe features of users and items.A number of experiments are conducted on three public datasets,and results show that the proposed method outperforms the comparison methods on multiple metrics.Further,this paper analyzes the representation distributions from independence and uniformity,verifies the disentanglement effect.It also compares the convergence speed to verify the effectiveness.

As an emerging task in text sentiment analysis,the automatic extraction of emotion-cause pairs aims to identify emotion expression from the raw texts without any annotation in the unit of clauses,and identify the causes for the corresponding emotions to form emotion-cause pairs.The crucial point of this task is focused on how to effectively capture the relationship between emotions and causes and among different emotion-cause pairs.To overcome the shortcomings of existing researches in capturing these associations,such as too coarse granularity and unable to effectively distinguish the mutual influence of causal relations between different pairs,this paper proposes an emotion-cause pair extraction method based on a heterogeneous graph neural network.Initially,we construct a heterogeneous graph with clauses and clause pairs as vertices,in which there are different types of edges between clauses and clause pairs and between different clause pairs to capture various fine-grained associations.Then using the heterogeneous graph neural network algorithm with attention mechanism to iteratively update the vertex embeddings of clauses and clause pairs.Finally,the updated embeddings is input to the binary classifier,and the classifier judges whether the corresponding pair has an emotion-cause relationship.To evaluate the effectiveness of the proposed model,we conduct a series of experiments on a benchmark dataset of the emotion-cause pair extraction task.The results demonstrate that the method based on the heterogeneous graph neural network proposed in this paper has a stable effect improvement,and the F1 value is 0.85% higher than the state-of-art baselines.When the bottom encoder(for obtaining the initial embeddings of clauses and clause pairs) is replaced by BERT,the F1 value can reach 73.12%,and our model also outperforms the state-of-art algorithm.

Current recommendation algorithms based on hashing technology commonly uses Hamming distance to indicate the similarity between user hash code and item hash code,while it ignores the potential difference information of each bit dimension.Therefore,this paper proposes a differential Hamming distance,which by calculating the dissimilarity between hash codes to assign bit weights.This paper designs a variational recommendation model for dissimilarity Hamming distance.The model is divided into a user hash component and an item hash component,which are connected by variational autoencoder structure.The model uses encoder to generate hash codes for user and items.In order to improve the robustness of the hash codes,we apply a Gaussian noise to both user and item hash coeds.Besides,the user and item hash codes are optimized by differential Hamming distance to maximize the ability of the model to reconstruct user-item scores.Experiments on benchmark datasets demonstrate that the proposed algorithm VDHR improves 3.9% in NDCG and 4.7% in MRR compared to the state-of-the-art hash recommendation algorithm under the premise of constant computational cost.

Reproducible build is the ability to recreate binary artifacts in a predefined build environment.Due to the role of reproducible build in ensuring the security of software construction environment and improving the efficiency of software construction and distribution,many open source software repositories(such as Debian) have carried out software reproducible build practice.However,due to the lack of sufficient judgment information and the complexity and diversity of source files,it is still a time-consuming and laborious challenge to determine why software can not be built reproducibly.In order to overcome this challenge,this paper studies the classification and detection of software unreproducible build causes based on machine learning.This paper stu-dies four typical reasons for unreproducible build,namely timestamp,fileordering,randomness and locale.This method uses the word vector generated by word2vec to represent the text log,and then cooperates with the logistic regression model to learn and train the text corpus combined with the difference log and the build log,so as to realize the automatic classification of the causes of unreproducible build.In this paper,the algorithm is implemented and tested on 671 unreproducible build Debian software packa-ges.Experimental results show that our method achieves a macro average precision of 80.75% and a macro average recall of 86.07%,which are better than other commonly used machine learning algorithms.In addition,we also analyze the relevance and importance of difference log and build log.Result indicates that both of them are significant for the classification of unreproducible build causes.This method provides a reliable research basis for automatic classification of unreproducible build causes.

RGBT target tracking uses the advantages of two different modes of visible light(RGB) and thermal infrared(T) to solve the common modal limitation problem in single mode target tracking,so as to improve the performance of target tracking in complex environment.In the RGBT object tracking algorithm,the precise location of the object and the effective fusion of the two modalities are very important issues.In order to accurately locate the object and effectively fuse the two modalities,this paper proposes a new method to explore high-rank feature maps and introduce position attention for RGBT object tracking.The method first uses location attention to focus on the location information of the object according to the deep and shallow features of the backbone network,and then focuses on the importance of the features by exploring the high-rank feature maps before the fusion of the two modalities to guide the modal features fusion.In order to focus on the object location information,this paper uses the average pooling operation on the rows and columns.For the high-rank feature guidance module,this paper guides the fusion of feature maps according to the rank of the feature maps.In order to remove redundancy and noise and achieve more robust feature expression,the feature graph with small rank is deleted directly.Experimental results on two RGBT tracking benchmark data sets show that compared with other RGBT target tracking methods,the proposed method achieves better tracking results in accuracy and success rate.

The hijacking attack for software online upgrade is one of the most common methods of network attack.Program ana-lysis is an important method to evaluate the security of software upgrade quickly and automatically.Rapid reverse positioning of upgrade functions in software is a key premise to realize static analysis and improve the efficiency of dynamic analysis.Traditional program analysis reverse localization relies on manual experience based on the cross reference chain relation of semantic information,such as string and API function,which is inefficient and cannot be automated.To solve this problem,this paper proposes a software upgrade function localization method based on semantic analysis and reverse analysis.Firstly,an upgrade semantic classification model based on natural language processing is established for common semantic information(string,function name,API function,etc.) in software binary program.Secondly,the software semantic information is extracted by reverse analysis tool,and the upgrade semantic classification model is used to identify the upgrade semantic information.Finally,an algorithm is defined to solve the key nodes of the upgrade function in the graph tree of function call relationship.This paper designs and implements a software online upgrade positioning system,and carries out reverse positioning analysis on 153 commonly used softwares,126 of which are successfully located.The security of some software upgrades is preliminarily evaluated by positioning analysis,and one CNNVD vulnerability and five CNVD vulnerabilities are found.

The emergence of business intelligence and analytics(BI&A) 3.0 and the broadening application scenarios of information fusion enhance the importance of data fusion in the business intelligence.More and more researches in the fields of economy,finance and management use the idea and methods of fusion,and the application of data fusion in these fields shows characteristics different from the traditional information fusion.Considering the concepts of information fusion and BI&A,this paper puts forward the new connotation of BI&A based on the perspective of data fusion under the background of multi-source and heteroge-neous big data,highlighting the importance of data fusion in BI&A.In addition,the paper constructs the fusion framework of ‘data,information and knowledge’ for BI&A based on WSR system methodology,so that the data fusion can be better applied in the fields of economy,finance and management.It provides scientific basis for acquiring knowledge from massive multi-source and heterogeneous data,and is beneficial to the development and implementation of a more effective business intelligence system.

In recent years,with the development of unmanned driving,intelligent UAV and mobile Internet,the demand for lightweight neural network from low-power,low-cost mobile and embedded platforms is increasingly urgent.Based on the idea of deformable convolution and depthwise separable convolution,this paper presents a low-cost deformable convolution,which has the advantages of high-efficiency feature extraction ability of deformable convolution and low computational complexity of depthwise separable convolution.In addition,on the basis of applying low-cost deformable convolution and combining with the method of model structure compression,4 lightweight methods of MobileNet network are designed.Experiments on Caltech256,CIFAR100 and CIFAR10 datasets demonstrate that low-cost deformable convolution can effectively improve the classification accuracy of lightweight networks without significant increase in computational effort.Besides,the accuracy of the MobileNet network can be improved by 0.4%~1% by combining the 4 MobileNet re-lightening methods in this paper,while the network computing load can be reduced by 5% ~ 15%,which significantly improves the performance of the lightweight network and better meets the practical needs of low power consumption and low computing power.It has very important practical significance for the advancement of intelligence in the field of mobile and embedded platforms.

The existing multi-signature scheme applied in the consensus mechanism scenario defaults that the signers are honest entities,so the security and validity of the signature could not be guaranteed when malicious nodes existed.In order to improve the robustness of multi-signature in the typical adversarial scenarios in consensus protocols,this paper proposes an ID-based multi-signature scheme based on the advantages of the ID-based cryptography system.In this signature scheme,non-fixed subgroup generates randomly cooperated to generate multi-signatures representing the entire group,and the validity of all subgroup signatures must be verified before signature aggregation.The bilinear pairings required by this scheme to generate multi-signatures are related to the number of subgroup members,which improve the security of the scheme at the cost of certain efficiency.This paper introduces a notion of robustness for robust subgroup ID-based multi-signatures,and the corresponding proof of the proposed scheme is given.Furthermore,under the random oracle model,relying on the hardness of the computational Diffie-Helman(CDH) problem,the scheme is proved is proved to be unforgeable under adaptive selection message attack.In addition,theoretical analysis and prototype implementation of the signature scheme are carried out,and the experimental results are compared with the performance of relevant signature schemes.

With the increasing globalization of football,the global player transfer market is becoming more and more prosperous.However,as the most important factor affecting player transfer transaction,the player’s transfer value lacks in-depth model and application research.In this paper,the FIFA’s official player database is taken as the research object.Firstly,on the premise of distinguishing different player positions,Box-Cox transformation,F-Score feature selection,etc.are used to perform feature processing on the original data set.Secondly,the player value prediction model is constructed by XGBoost,and compared with the main machine learning algorithms such as random forest,AdaBoost,GBDT and SVR for 10-fold cross validation experiments.Experimental results prove that the XGBoost model has a performance advantage over the existing models on the indicators of R2,MAE and RMSE.Finally,on the basis of constructing the value prediction model,this paper integrates the SHAP framework to analyze the important factors affecting the players’ value score in different positions,and provides decision support for some scenarios,such as player’s value score evaluation,comparative analysis,and training strategy formulation,etc.

Edge intelligence refers to a form of service that uses artificial intelligence algorithms to provide data analysis capabilities for network edge devices.However,the edge computing environment is more complex and changeable than cloud computing.There are many problems in the process of building edge intelligence,such as the lack of quantitative evaluation standards,heterogeneous computing platforms,complex network topologies,and changing user needs.Among them,the more prominent is the contradiction between the high resource demand of the algorithm model and the low resource reserve of edge devices.Machine lear-ning is the main workload of edge intelligence.It requires a lot of computing resources.However,the computing resources of edge devices are limited,and the supply and demand between the two do not match.The deployment and optimization of edge intelligent load has become a problem.Therefore,in response to the problem of edge intelligent load performance optimization,this paper proposes cloud-edge collaborative inference(CECI) based on load characteristics,which is optimized for different machine learning loads in terms of model selection,batch adaptive adjustment and cloud-side collaboration.In terms of model selection,a model adaptive selection strategy based on target weights is used to comprehensively weigh the effects of multiple performance optimization targets under multiple constraints.In the aspect of batch adaptive adjustment,a batch adaptive adjustment algorithm based on overhead feedback is proposed,so that the model can achieve better performance at runtime.In terms of cloud-side collaboration,a cloud-side collaboration strategy is designed by combining network status and user delay requirements to achieve the effect of dynamic utilization of cloud computing resources.Experimental results show that compared with cloud intelligence,the edge intelligence based on load characteristics proposed in this paper can reduce program running time by 50.79%,reduce system energy consumption by 42.46%,and improve model accuracy by 4.52%.

Traditional risk management methods focus on identifying,predicting and assessing potential risks.However,when enterprises are exposed to uncertainty and unexpected risks,traditional methods cannot deal with those risks.Therefore,the academia gradually shifts the perspective of risk management from predicting and avoiding risks to improving the ability of enterprises to withstand and recover from risks,that is,the enterprise resilience.This paper proposes a prediction method to predict the enterprise resilience based on temporal features,which utilizes Bi-LSTM to encode the temporal features to obtain the feature representation of every enterprise,and the classification results of enterprise resilience are obtained by a softmax classifier.The proposed method is validated on the real-world datasets from listed companies in China,and the macro-F1 value reaches 89.0%,which is improved compared with those models without considering temporal features,such as RF,XGBoost and LightGBM.This paper further discusses the importance of various factors that have an influence on the enterprise resilience.In this paper,the machine learning methods are applied to the evaluation and prediction of enterprise resilience for the first time,which provides theoretical and methodological guidance for enterprises to deal with unexpected risks.

Patent knowledge graph plays a important role in patent accurate retrieval,patent in-depth analysis and patent know-ledge training.This paper proposes a practical patent knowledge graph construction method based on seed knowledge graph,text mining and relationship completion.In this method,to ensure the quality,a seed patent knowledge graph is first established ma-nually,then the concept and relation extraction method of patent text pattern is used to expand the seed patent knowledge graph,and finally the extended patent knowledge graph is quantitatively evaluated.In this paper,artificial extraction of seed knowledge and manual summarization of lexical and syntactic patterns are carried out for patents in the field of traditional Chinese medicine.After obtaining new lexical and syntactic patterns by machine learning,the knowledge graph of seed patent is expanded and completed.Experimental results show that the number of nodes and relationships in the knowledge graph of traditional Chinese medicine are 19 453 and 194 775 respectively.After expansion,they reach 558 461 and 7 275 958 respectively,representing an increase of 27.7 and 36.3 folds respectively.

With the increase of temporal data,the concept of temporal knowledge graph is popularized,and how to represent temporal knowledge graph efficiently has become an important research direction.Although resource description framework(RDF) is widely used in traditional knowledge graph modeling,it can only represent static semantics and lacks the ability to represent temporal knowledge graph.Therefore,several temporal RDF models have been proposed for temporal knowledge graph,but all these models simply attach temporal information to the predicate of RDF or the whole triple,and lack the accurate positioning of the object to which the temporal information belongs.In order to better represent temporal knowledge graph,firstly,this paper proposes a new temporal RDF representation model called tRDF,which attaches temporal information to the object or predicate according to the type of object.Secondly,by combining the concept of temporal database,this paper presents a tRDF data storage method based on the relational database,PostgreSQL.Finally,the proposed tRDF data storage method is verified from two aspects,the time of storing and the size of space.Experimental results show that the proposed scheme can effectively represent temporal knowledge graph.

Traditional cloud providers provide services solely for users,which has problems of insufficient local cloud resources and high expansion costs.While the emerging multi-cloud combines services of cloud providers with different geographical locations providing users with more choices and has gradually become a research hotspot.At the same time,workflow scheduling is one of the key problems in multi-cloud research.Therefore,this paper firstly makes a thorough investigation and analysis on workflow scheduling technology in multi-cloud environment,then compares and classifies the workflow scheduling method inclu-ded.This paper focuses on single objective workflow scheduling optimization problem for cost and completion time,multi-objective optimization workflow scheduling problem for cost and completion time,for response time and cost,for reliability,cost and completion time,and for other multi-objective optimization.Finally,the future research directions of workflow scheduling in the multi-cloud environment are discussed,including uncertain workflow scheduling,joint scheduling optimization of energy consumption and other objectives,scheduling optimization with edge servers,and hybrid scheduling of virtual machine and Serverless platform.