This paper studies uncertain graph data mining and especially investigates the problem of mining dense subgraphs from uncertain graph data.Based on the uncertain graph data model with weighted edges,expected density of subgraphs and expected degree of vertexes are employed to measure the dense degree of subgraphs.The characteristics of EDP(expected density peak) is proposed during greedy iterations,which improves the algorithm execution with 2-approximation result and an efficient execution.The algorithm is proved to guarantee the correctness of the final mining result.When the subgraph has a size constraint,the problem of mining dense subgraph becomes NP-hard.Compared with other methods,the improved dense subgraph mining algorithm with size constraint is more efficient.

Most electronic commerce(Ecommerce) systems use Collaborative Filtering (CF)based methods to recommend items belonging to different categories.Market basket analysis has found that a group of likeminded users have similar tastes on items belonging to a subset of correlated categories(called crosscategory dependence) rather than all the categories.Therefore，we should consider both usertouser similarity and itemtoitem similarity in recommendations.In real applications，items are usually organized into a multilevel taxonomy which provides hierarchical relationships between items and categories.Note that，the degree of data sparsity varies in different level of categories as there are more items in a category than those in any of its subcategories.To alleviate the data sparsity problem of existing recommendation methods，we propose an efficient multilevel biclustering algorithm to mine useritem/category biclusters(i.e.crosscategory dependence) at each level of the taxonomy.Then we propose a general framework for crosscategory recommendation which extends existing CF methods by utilizing multilevel biclusters to improve their recommendation performance.Experiments on a real datasets show that our recommendation framework based on multilevel biclusters can recommend correlated items to a group of users precisely and efficiently.

More and more organizations and individuals outsource their storage and computing needs into a new economic and computing model, which is commonly referred to as cloud computing. As one of the hottest research issues, cloud computing security is also concerned by many scholars. This paper survey the recent advances in some security issues of cloud computing, which mainly focuses on the topics of data security, identity authentication and access control policy in cloud computing environment. It also introduces some enhancement frameworks and projects about the cloud security integrated with the Trusted Computing. According to these surveys, this paper believes that the combination between trusted computing and cloud computing will be a promising direction of the future cloud security researches, and proposes some interesting research issues in the end.

An efficient Artificial Bee Colony (ABC) algorithm for function optimization was proposed in this paper，in order to overcome the drawbacks of low computational accuracy and slow convergence of conventional ABC algorithm.In this algorithm，the chemotaxis operator in bacterial foraging optimization algorithm was introduced into the local search process of onlooker bees，and onlooker bees could do the local search around the current optimal solution which after employed bees dimension variation，which was to enhance the local search capability of the algorithm.The simulation results of six standard functions show that the modified ABC algorithm can attain significant improvement on solution accuracy and convergence rate，when compared with the basic ABC algorithm.

This paper introduces the concept of data provenance, and investigates it from three aspects: method, model and application. Seven data provenance models are presented: flow information model, time-value centric model, four dimensions model, open provenance model, Provenir model, data provenance′s security model and PrInt model. Based on these models, we provided a data provenance model for heterogeneous data. Based on the analysis of several popular data provenance methods, we proposed a method to extend the labeling method with column storage. We described data provenance applications in the fields of database, workflow and other. The typical application cases are provided. At last, the hot research spots and directions are given.

Collaborative filtering is one of the most popular recommendation algorithms,which has been successfully used in many recommendation systems.Among different approaches,latent factor model is a representative one.The core idea of latent factor model is to connect user′s interests and item by means of the latent classes,then the matrix factorization technology is used to discovery the relationships between the users and the latent classes,the relationships between the latent classes and the items.Finally,we can infer the degree of preference of a user to each item,and provide personalized item recommendations that will suit for this user′s tastes.However,this approach still suffers from several inherent issues such as data sparsity and cold start.Luckily,along with the rapid development of social network,many scholars have fused the social media data (such as label,social,etc) into the latent factor model to solve the problems that collaborative filtering encounters.In this paper,we review the research of recommendation algorithms based on latent factor model in recent years,summarize the common topologies of the latent factor model,and point out some future research directions.

In humancomputer interaction，the most natural and the best way to exchange is the communication by human voice.Which is mainly related to speech synthesis，that is，the technology of converting text to speech.This paper provides a concise but deep introduction to the development of speech synthesis and propose the problems and solutions in the development of speech synthesis.The researchers who are just entering the field of speech synthesis can stand on the shoulders of giants，and have a clear and deep understanding of voice synthesis，and start working with a correct judgment.In this paper first the effective methods which are already exist and mainstream in speech synthesis are overall introduced and，the main idea of these methods as well as their advantages and disadvantages are described.On this basis，we inspire new ideas.After that，this paper illustrates the efforts respectively at home and abroad in recent years that researchers have done in the field of speech synthesis.Then objective evaluate and analyze gains and losses in synthesis technology improvements and draw the trend of synthesis technology in recent years.Finally，get the prospect in speech synthesis，pointing the bottleneck in development and trying to give direction to solve them.

As an advanced memory device with high density, Phase-change memory is comparable with DRAM in reading. So it is increasingly considered as a prevalent DRAM alternative due to its high density, low standby power and bit addressability, especially in the design of mobile computing systems with limited power supply. But long write latency, large write energy and limited endurance are becoming obstacles of PCM technology. An important concern for researchers is to achieve high-performance and low-power memory access according to the innovation of memory architecture using this new type of memory technology sufficiently. In this paper, we propose a Memory Access-aware Remapping Mechanism for PCM-DRAM hybrid memory system to tackle the above problems. This system places PCM and DRAM in the same address space and implements the management of PCM and DRAM physical page frame using address remapping mechanism, according to real-time monitoring of memory access pattern. Experiment results demonstrate an Energy-Delay Product (EDP) improvement of 32% on average in MARM-based PCM-DRAM hybrid memory system, compared to conventional DRAM memory system.

As an efficient and reliable method in 3D reconstruction and active measurement, structured light vision technology is being more and more important. In a structured light system, accurate reconstruction depends on the correct correspondences between the points in the image and the points in the projected pattern, where a proper codification method would be crucial in real applications. In this paper, after a brief introduction of the coded structure light system, a survey is then presented to summary the existing codification methods. The advantages and limitations of the different methods are discussed. Finally, a summarization of the codification methods and a discussion of future trends are provided.

While many image classification problems are featured by intraclass similarity and interclass dissimilarity，the phenomenon of interclass similarity and intraclass dissimilarity often occurs in flower image classification，thus yielding unsatisfactory flower image classification accuracy if traditional manual features are used.To solve this problem，a method combining deep features and manual feature of flower images is proposed for flower image classification.Firstly，a unified depth feature extraction framework is constructed based on deep convolutional neural network.Then，color and intensity features are obtained via CNN model respectively.Next，we design a texture feature based on CNN through the feature map in the low layers.Finally，after the multifeature fusion with above features and a manual feature，a more comprehensive description of flower image is acquired.The classification experiments show that the proposed features at lower dimensions than the traditional ones achieve better performance in flower image classification.

As the Web 2.0, network becomes one of the important medium for reflecting public opinions, finding and mining public opinion orientation become an issue. But till now, no effective opinion monitoring system of reflecting the total orientation of the netizens on some hot events or topics has been proposed. In this paper, according to the characteristics of simplicity and large amount of opinions on some topics, an existing sentiment words lexicon is extended using HowNet and NTUSD, and a new sentiment lexicon with sentiment orientation extent is built. Based on the extended sentiment lexicon, a semi-automatic web public opinion analysis system is proposed, which can provide users more detail and precise opinion orientation analysis results.

In order to study the problems of traditional target detection algorithms in the detection of road small targets,such as poor performance and high missed detection rate,this paper proposes a small target detection method based on improved YOLOv3.First,reduce the missed detection rate of small targets by designing a new feature fusion structure,and use DIOU loss to improve positioning accuracy.At the same time,the clustering algorithm in the YOLOv3 algorithm is improved,the K-means++ algorithm is used to improve the extraction of the center point of the clustering prior box,and a more suitable Anchor Box is selected to improve the average accuracy and speed of detection.Pedestrians and vehicles are comparatively detected on a self-made mixed data set.Without affecting the detection speed,the improved YOLOv3 algorithm can effectively reduce the missed detection rate of small objects and improve the detection accuracy.According to the experimental results,the improved YOLOv3 model proposed in this paper has an average accuracy of 92.82% on the mixed data set,which is an increase of 2.77% compared with the unimproved YOLOv3 algorithm.

At present，most open text classification datasets are relatively balanced，while the class distribution in the real world is extremely imbalanced.This paper proposes a text classification algorithm based on convolutional neural network and attention mechanism，named AM-CNN （Convolutional Neural Network with Attention Mechanism）.The algorithm captures the context information of the text based on time recurrent neural network at first.Meanwhile，to reduce the impact of small categories，attention mechanism is introduced to extract the feature vector matrix of the text categories.With the matrix result，AM-CNN is designed to complete the text classification task based on convolution neural network.The experimental shows that the proposed algorithm treats the texts fairly during the model training process，improves the accuracy of text classification significantly.

Massive medical data not only contains great value，but also implies a great deal of personal privacy.For the information security of medical data，this paper proposes a blockchainbased allhomomorphic medical data security sharing scheme，which can calculate and apply medical data in ciphertext state in a decentralized network.Through the combination of blockchain technology and all-homomorphic encryption technology，the mapping of centralized nodes such as medical institutions，patients，and third-party data processing centers in the centralized network is first implemented in the blockchain network to achieve decentralization and completeness of the nodes trust.Then，the full homomorphic encryption algorithm is invoked through the smart contract to implement the transmission of ciphertext data between the two parties and the ciphertext calculation can be performed.The ultimate goal is to ensure personal data privacy，data distribution，and secure transmission without affecting the analysis and practical application of medical big data.

Extensive and longtime business operations contribute to the business big data.In the Internet ear,abundant business process models and executing logs become available,which could be reused in the practice of business process modeling.Considering the fact that limited sources,static criteria,and coarse granularity employed by traditional process mining and process retrieval,there still call for effective approaches to take the advantage of this opportunity in order to highlight the creative nature in business process modeling.To meet the requirement,we propose a novel recommendation approach in process design based on the average perceptron over multiple information sources.It is characterized by taking into account historical logs as well as available process models,and combine human creativity and the adaptability to real situations at the stage of business process modeling.Experimental results against three different types of data demonstrate its excellence in simulating human behaviors.Specifically,promising performance are achieved to the real bioinformatics process data(53.69%),related work data set (gain with +42.35%),and synthetic data set by ProM′s PLG tool (gain with +9.46% by considering both case structure and log items,or +5.94% gaining in contrast to merely logs).

Nowadays it is difficult to distinguish the subtopics in a hot news topic on the internet. To solve this problem, in the paper, the method of subtopic division based on Latent Dirichlet Allocation is presented. It describes a news document by Latent Dirichlet Allocation, and uses Bayes standard method to determine the optimal number of topics in order to fit documents best. According to the high similarity of documents between subtopics, the relativity analysis of feature words is introduced. Using the improved Kullback-Leibler distance to calculate the similarity of news stories can distinguish the stories which have similar content but belong to different topics effectively. Finally, it divides a hot news topic to subtopics by clustering the news documents with the single-pass incremental clustering algorithm. Experimental results verify the availability of the improved similarity calculation method, and it shows that this method can improve the performance of subtopic division effectively comparing to the baseline method.

Recently,machine learning algorithms have been widely used in many fields.Hyperparameter directly affects the performance of the machine learning algorithms.However,hyperparameter tuning depends on the professional knowledge and the expert experience.In order to solve the above problem,we propose an automatic hyperparameter optimization method based on reinforcement learning.This method considers the hyperparameter optimization problem as a sequence decision problem and models it as a Markov decision process (MDP).An reinforcement learning agent automatically selects hyperparameters for a machine learning algorithm.The accuracy of the model on the validation data set is used as a reward.To reduce the variance during the training,a data boot pool technique is designed.We have conducted a series of experiments to tune hyperparameters for the Random forest and XGBoost.We have compared our method with five optimization methods: random search,Bayesian optimization,TPE,CM-AES and SMAC on five datasets.The experimental results show that the proposed method achieves the best performance on 90% of the tasks..In addition,we have verified the effectiveness of the agent structure and the data boot pool by performing the ablation experiments.

Nowadays，functional modeling and fault analysis are separated for safetycritical software in complex embedded system，in the areas of aeronautics and astronautics，nuclear power and others.It leads to the problem that functional model analysis is lack of safety property and the hazard getting from fault tree analysis can′t be avoided during functional design.State／Event fault Tree (SEFT) is a modeling technique for describing the causal relations which lead to functional failure in complex systems，it can unify functional modeling and fault analysis；but because of the lack of semantic precision，it can hardly be used directly for software safety analysis.A method for software safety analysis based on SEFT is presented in this paper.Firstly，translate SEFT to state machine addition with fault semantic messages by means of mapping elements together with translating logic gates；after which，translate state machine to timed automata；at last，test software safety to collect counterexamples for analysis using model checker UPPAAL.A case study of gas burner control system is given in this paper.

Cloud computing is a way to quickly increase the capacity without investing in new infrastructure，training new personnel，or licensing new software.So in the last few years，cloud computing has grown from being a promising business concept to one of the fast growing segments of the IT industry.However most of enterprise customers are still reluctant to deploy their core businesses onto the cloud.One of the reasons for this is lack of trust.For example in the cloud computing the users lose control over their data；Moreover the cloud servers are not transparent；Furthermore the security assurances given by cloud service provider are not clear and so on.Therefore in order to enhance consumer′s confidence in the cloud computing，the trust mechanisms come into being and become a hot topic.In this paper we survey the trust mechanism in the cloud computing.Roughly the current trust mechanism can be divided into five categories：reputationbased，SLAbased，auditbased，intercloudbased and encryptionbased.At first each mechanism is emphatically analyzed as follows：the general framework is introduced as well as the workflow；also the typical methods are evaluated and compared；Then we give the properties of trust and security with which each mechanism can meet；Finally we describe the future research directions of these five mechanisms.

In recent years,big data processing related theories and techniques get more and more attention from industry and academic.On the one hand,the scientific research produces a large amount of data.Analyzing these data is an important part for scientific research.On the other hand,with the continuous development of information technology,enterprises accumulate a large amount of structured and unstructured data during informatization process.How to manage and operate these data has become the company′s core assets,profoundly affect the company′s business model,decisionmaking,organization and business processes.Therefore,a large data processing related technologies have also been of great concern to the industry.Based on the time characteristics of the data processing,big data processing mode can be divided into three modes —— offline batch data processing,querybased data processing and realtime data processing.This article summed up the general framework of big data processing from a technical point of view,and carry out discussion of each levels of big data processing for different processing mode.Because big data processing is based on big data storage,we first put some discussion on big data storage.Then we expand the description of these three modes so that the readers could have a preliminary understanding about big data system building.

With the development of deep learning,convolutional neural networks have led to remarkable success in object detection.Compared with the traditional object detection algorithms based on artificial features construction,the algorithms based on deep convolutional neural network have the advantages of automatic feature extraction,strong generalization ability and good robustness.This paper firstly introduces the progress of convolutional neural network on the classification which is the base of object detection task,and then analyzes and compares the object detection algorithms based on deep learning model in recent years according to three aspects of object detection,including algorithm evaluation metric,algorithm frameworks and public datasets.Finally,this paper forecasts the future development of the object detection algorithms.

Analysis of malware and antianalysis technique has always been the focus of the computer security field.Malware implements the selfprotection by antistatic analysis and antidynamic analysis：antistatic analysis uses the method of packers and code obfuscation to disturb disassembly and identification of control flow；antidynamic analysis detects system operating environment information to realize the antitracking for debugger and virtual machine.Correspondingly，unpacking，antiobfuscation and virtual machine technology are used by analysis and antivirus software to avoid interference by antianalysis techniques.Paper indepth analyzes and summarizes the principle of various malware antianalysis and analysis technology，explores the advantages and disadvantages and applicability of these technologies，provides some ideas and technical direction for the development of malware analysis techniques.

Biomedical named entity recognition is the prerequisite for biomedical information extraction.The current entity recognition methods,which are based on machine learning,mainly depend on manually summarizing features,according to the domain knowledge and experience,and need to do experiments repeatedly for selecting the appropriate features.And these features rarely utilize the deep semantic information.To investigate the effect of semantic information on Named Entity Recognition,this paper attempts to obtain semantic information automatically from the large-scale unlabeled corpus,which can be downloaded from public database,such as PubMed,and get three kinds of word representation approaches,including word embeddings,cluster based on word embeddings,and Brown cluster.The three kinds of word representation are adopted as the features of CRF model and SVM model for semi-supervised learning.Comparative experiments are conducted under the same conditions：the dimension of word embeddings and the number of clusters.The experimental results show that the word representation approaches can learn the latent semantic information effectively and thus improve the performance of existing entity recognition systems based on machine learning.Experimental results （Precision,Recall,F-score） on public evaluation corpus BioCreative II GM reaches 9124％,8580％,and 8844％ respectively without the dictionary or any other external resources.

The traditional feature selection methods handle data with balanced distribution, aim for getting optimal classification accuracy, so there exist very limited feature selection methods that perform well on imbalance data. This study proposes a new feature selection method based on the character of data distribution for imbalanced data sets. It modifies data distribution for balance by assigning different weights to the function of feature importance measurements according to the variation of the size of clusters in unsupervised learning. Experimental results on several UCI datasets show that the performance of the proposed method outperforms other classic feature selection algorithms. It not only maintains or enhances the classification performance and dimensionality reduction, but also improves the precision, recall and F-Measure of the minor classes on different classifiers.

With the rapid growth of the number of Web services in the Internet, QoS has played an increasingly important role in services selection. To obtain Web services QoS and recommend Web services to users effectively, an user location-based QoS predicting method, i.e. UL-WSRec, is proposed in this paper. Based on the fact that parameters of QoS usually depend on users′ locations, the method computes user′s similarity by considering the characteristic that users reside in the same autonomous system of Internet are similar in locality. By incorporating users′ locality information, the method adapts the traditional collaborative filtering algorithm so as to improve both efficiency and precision of Web services QoS prediction. Experiments conducted on real dataset of Web services validate the approach.

Knowledge graph，a production of big data era，was used as a new approach of knowledge representation and model of knowledge management.It was usually applied in areas of large scale knowledge graph completion，information retrieval，natural language processing and machine learning，etc.Knowledge graph learning and reasoning was shown as an effective way to solve these problems and was a core content in applications of knowledge graph meanwhile.Therefore，studies on algorithms of knowledge graph learning and reasoning had great significance for knowledge graph application.This paper firstly summarized and introduced the main tasks of those studies，including link prediction，entity resolution and Link-based clustering，etc.Then analyzed and compared the advantages and disadvantages of various algorithms grouped by class，based on the review and detailed introduction of the state-of-the-art algorithms of knowledge graph learning and reasoning.After that，it summarized the advantages and disadvantages of three typical classes of algorithms.Moreover，it discussed the current major problems we are facing and the possible extensions.It suggested existing algorithms should improve predicting efficiency and accuracy，most of the knowledge of learning and inference algorithm should not just applied to binary relations，application fields should extend from general knowledge graph to professional areas，and to pay more attention on multidata fusion reasoning and Chinese knowledge graph learning and reasoning.

Symbolic execution is widely used in program analysis for its wellcontrolled precision.It denotes the value of variables as abstract symbols and simulates the execution of programs.Because our tool is pathsensitive,we suffer the path-explosion problem when analyzing source code.We present a program slicing method for defects to alleviate the problem.Firstly,we generate the slicing criterion of source programs according to defects that users concern about.Then we analyze the source code to generate Data Dependence Graph and Control Dependence Graph,which constitute Program Dependence Graph.Next,the program is sliced according to the slicing criterion so that the source code could be reduced.Finally,we use the program analyzer to analyze the sliced program.The experimental results on ShapeChecker,our symbolic execution tool,show that the method is effective.

With the development of Internet of Things technology,the structure of Internet of Things is becoming more and more complex,the amount of data collected shows explosive growth,thus how to safely share data between different participants has become a huge challenge. The traditional data sharing model is often dependent on the trusted third party centralized organization,but this scheme is likely to cause a single point of failure,is not transparent to participants,and data may be tampered with. To tackle this problem,we propose a data sharing model for Internet of Things based on blockchain,which uses the feature that blockchain can establish trust without any centralized organization. We first analyze the existing data sharing models of Internet of Things and the key concepts of Hyperledger Fabric blockchain platform. After that,the design of gateway and the content of data stored in blockchain ledger are illuminated,the enhancement methods of security and data privacy are proposed. Finally,the security and availability of the model are analyzed,and the feasibility of the model is proved by simplified testing.

In recent years，with the rapid development of knowledge base，many high-quality，large-scale knowledge bases have emerged，and the Knowledge Base Question Answering（KBQA）system has developed rapidly with the development of the knowledge base.The KBQA system understands and parses natural language questions，and then uses facts in the knowledge base to automatically answer natural language questions.Users can quickly and accurately obtain valuable knowledge or information without understanding the data structure of the knowledge base.This article introduces the research methods of the KBQA system in detail and summarizes the current research progress，including query template-based methods，semantic parsing-based methods，and deep learning-based methods.By comparing these research methods，we point out the shortcomings and problems in each technique，and then summarize the issues and challenges faced by the KBQA system.

Fine-grained trust level distinction in multi-access control is not well resolved in traditional role-based access control (RBAC) model. In this paper, a variety of role-based access control models and attribute characteristics are deeply researched, and an attribute-policy-based RBAC model is proposed, then defined formally. Attribute-policy-based RBAC model extends the concept of RBAC roles, defines properties for roles and provides an attribute-policy-based authentication. The model gives the realization of precise and flexibility multi-access control and improves access control accuracy for fine-grained data objects. In cloud computing platform, a SaaS model of fine-grained object management services is designed and implemented. Experiment shows that the model is adaptive to changes in the dynamic permissions and has ability to control multi-access control.

According to the small world effect in netizens relationship network topology, puts forward the opinion tendency conversion rules. Based on the small-world networks matrix of the netizens relationship in the network public opinion, constructs internet public opinion evolution migrant cellular model based on small world effect, uses this model analyzing network public opinion evolution and produces the simulation results, which include tendency conversion figure, curve of course tendency and curve of fine tendency. Through the emulation result analysis, reveals the opinion polarization phenomenon and "core" regional drift phenomenon in network public opinion evolution, analyze the causes of the "core" regional drift phenomenon and tendency multi-wave curve. Emulation results show that the model can better fit with the network public opinion evolution laws. This research provides certain theoretical basis for the management in network supervisory departments and news management department.

The big data era has come，which brings new challenges and opportunities to the information science community.The traditional methods and tools are gradually difficult to meet the requirement of companies.The research on analysis and storage for big data is topical and important.In this paper，we illustrate the basic concept of big data，“four Vs” features and the big data lifecycle，as well as review several data processing models and data structures.Then，we not only summarize hardware support but also introduce a variety of typical system supports for big data analysis，such as Hadoop，Spark，Storm and Petuum，which are compared according to their applications.Finally，we present several typical applications for big data analysis in the field of government，education，transportation and medical treatment.

In order to simulating cutting on soft tissue in the virtual endoscopic sinus surgery, cutting simulation of soft tissue in virtual surgery was discussed. The tetrahedral mesh model, and the simplified model of the moving of scalpel were used, the cutting tree was constructed for the cutting operation, the cutting division and cutting thinning algorithms were designed which can reduce the increase number of grid cell structure by duplicating and moving vertices according to their location in the cutting. Through simulation experiments, the algorithm′s time-consuming keep the linear relationship whit the length of cutting; after using this algorithm to simulate cutting in the grid structure, the Incremental in the number of cell structure was less than using traditional algorithm significantly.

Body Area Networks technology is an important applications direction of wireless sensor networks ,it has broad application prospects in telemedicine, health care and other fields. of the technology to solve issues such as medical treatment is difficult and expensive, and the response to the problem of population aging has Important practical significance. The technologies of body area network are introduced in detail. Firstly, technologies the concept of Body Area Network technology and applications are presented. Then, based on existing studies , a summary of the health care field for a typical body area network architecture is given. After that, from wireless communication, media access control, security, the current research issues are discussed. Finally, it putforward the conclusion and the prospects.

Aiming at the user preference diversity and interests dynamics in real shopping scenarios，this paper proposes a personalized sequence recommendation deep learning model that fusion the contextual information.More effectively explore user′s long-term and short-term interests by embedding feedback information provided by users and effectively solve the problem that the traditional recommendation system can′t simulate the evolution of user interest.This article takes real e-commerce website data as a background.Firstly，it uses historical behavior data and item auxiliary information to construct long-and short-term session sequences and fuse contextual information，and proposes interest attenuation factors to reflect changes in user preferences.Secondly，based on the TextCNN model training to obtain sequence vector representation and extract the user item sequence latent vector through the multi-head attention mechanism；Finally，the combination vector of user′s cross information and potential behavior characteristics is fed to the multilayer perceptron to establish a sequence-based recommended model.Experimental on two real datasets and a public datasets，show that adding interest decay factor and project auxiliary information to the behavior sequence improves the model performance.In addition，the prediction model based on this paper has improved the evaluation indicators RMSE and GAUC compared with the traditional recommendation algorithm.

User modeling is the prerequisite and core technology of personalized service offering. It determines the quality of personalized service. User modeling is divided into five related modules, i.e., input, output, time characteristic, objects to be modeled and algorithms. The current status of the five modules is reviewed in detail. Input reflects the necessary data source for user modeling. Output is the model representation. Time characteristic means the effective duration of user model and update frequency. Objects to be modeled list the target objects to be included in the user model. Algorithms are applied to transform the input to output.

The traditional financial time series forecasting methods use accurate input data for the object of study, and then make single-step or multi-step prediction based on the established regression model.So its prediction result is one or more specific values.But because of the complexity of financial markets,the traditional forecasting methods are less reliable. In this paper, we transform the financial time series into fuzzy grain particle sequences ,and use support vector machine regression to regress the upper and lower bounds of the fuzzy particles, and then apply regression model single-step prediction on the upper and lower bounds,which will limit the predict results within a range.This is a new idea. The Shanghai Composite Index Week closing index for the experimental data, experimental results show the effectiveness of this approach.

With the rapid development of Internet technology, privacy preservation has been an essential issue for individuals or organizations. The emergence of kinds of data mining tools makes privacy disclosure issues be increasingly critical. The method of releasing data through removing identifier from the table could not truly prevent privacy disclosure. The attacker can also infer the private data of the corresponding individual with high probability by linking operation. Anonymization is one of the primary techniques realizing privacy protection in data dissemination environment. The paper simply introduces the general concept and basal principle of the anonymization techniques, and mainly analyzes and summarizes the progress of the anonymization principles, anonymization methods and anonymization metrics. Finally, the present problems and directions for future research are discussed.

Mandatory Access Control is one of the essential secure mechanisms for a DBMS.In a DMOSMAC system,the mandatory access control is implemented based on the mandatory access control mechanism of a secure operating system.The system is verified formally and the implementation of the system is analyzed.The concept of information flow is presented in this paper.The set of information flow is one of the elements of the system status for the DMOSMAC system and the set of the information flow increases so that the circumvention of verification for delete operation (and so on) is prevented.Therefore,the rigidity of the verification is preserved.Based on information flow,a method of formal verification for source codes of a DBMS system is proposed:extracting the system status and operational rules from source codes of DMOSMAC to ensure that the verification is accordance with the implementation;mapping the system status and operational rules of the BLP model to the status and operational rules of the DMOSMAC;adapting the simple security property and *-property in BLP model into the invariants for information flow;verifying the security based on the axioms and theorems in the BLP model.The theorem prover COQ is used to verify the security of the DMOSMAC system.

The emerging popular Internet applications such as online photo, audio, video and micro blogging services exhibit very different data access and storage requirements from traditional applications. Large number of small data are generated, analyzed, and returned every second in data center. These applications requires challenging performance for highly concurrent both high throughput and low latency access to tiny files. In this paper, we propose HVFS, a novel distributed file system built over distributed tabular storage, to manage billions of small files and support highly concurrent accesses. HVFS uses extendible hash to index metadata, log-structured storage format and columnar storage to exploit temporal and spatial locality. We present the design and implementation of HVFS. Our evaluation results demonstrate that the core tabular storage of HVFS can serve more than 100,000/240,000 aggregated data read/write requests per second on 82 nodes (《1KB), and FUSE implementation can serve more than 180,000 aggregated file creations per second on 32 nodes.