【作者】 李朋；

【导师】 文俊浩；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2013， 博士

【摘要】 随着信息数据类别的多样化和数据关系的复杂化，信息网络正在向异构化方向发展。因此，如何借助网络分析的手段，从异构信息网络中挖掘出有用知识是信息检索和知识挖掘面临的新课题。在异构信息网络中，参与知识挖掘的关键元素主要包括数据、服务和人类活动。上述元素中，以关系型数据库为代表的数据存储方式为海量信息提供了结构化的数据管理模式；以Web服务为代表的功能提供方式为构建公开化、松耦合的信息平台奠定了基础；以微博为代表的社交网络活动形式提供了新型的数据共享和信息交互方式。随着数据类别的多样化、服务访问的频繁化以及社交活动的网络化，人们对个性化的数据查询、聚类分析、活动预测等需求与日俱增，因此，对异构信息网络分析模型及其在信息检索和知识挖掘中的应用研究具有理论及现实工程意义。针对异构信息网络发展趋势及面临的新课题，基于异构信息网络中异构对象关系挖掘与异构信息网络描述模型，研究了异构信息网络中节点排序函数；基于描述模型和排序函数，结合Web服务异构网络、关系型数据库元组网络与社交网络，研究了异构信息网络分析模型的新型聚类分析、排序以及活动预测方法。论文研究的主要工作包括：①结合信息网络异构化发展趋势，基于对聚类、个性化查询与社交网络预测等研究现状及存在问题的分析，借助形式化方法研究了异构信息网络的描述模型。②基于异构信息网络描述模型，提出了基于异构信息网络分析的排序方法。根据不同网络连接形式和排序规则，该排序方法定义了4种不同类型的排序函数。不同排序函数的实例分析对比研究表明，该排序方法可为网络分析提供基础数据排序方法支撑。③鉴于以属性为计算依据的聚类不支持异构数据、忽略数据排序等问题，从关系的维度出发，提出了基于异构信息网络分析的聚类算法。基于该聚类算法，以Web服务聚类为例，提出了基于异构服务网络分析的服务聚类算法SNTClus。SNTClus算法基于服务标签等各参与方对象及关系构建异构服务网络描述模型，基于服务排序模型构建聚类多维度量模型，借助网络划分和排序循环迭代方法实现Web服务聚类。以Titan服务集为数据集的实验分析结果表明，SNTClus算法的服务聚类时间开销代价低、聚类准确度高。④针对当前信息查询中个性化支持程度低等问题，提出了基于异构信息网络分析的个性化查询方法。该方法研究以关系型数据库为例，针对当前关系型数据库个性化top-k查询要求，提出了基于异构元组网络分析的关系型数据库排序方法RNRank。RNRank排序方法基于异构元组网络提取和异构元组网络关联分析构建元组排序模型，按照是否考虑数据类别属性分别提出单类别数据元组排序算法RNRank-I和多类别条件下基于聚类分析的数据元组排序算法RNRank-II。以IMDB电影数据集和German Credit信息卡数据集为测试数据库的实验分析结果表明，RNRank算法具有较高的数据排序效率和排序准确度。⑤为了实现对社交网络活动的预测，提出了基于异构信息网络分析的社交网络活动预测模型。该模型基于不同网络结构特性的4种信息活动预测方法，通过相对准确率对比，采用不同加权系数构建综合预测公式，实现对社交网络消息转发次数和可能浏览次数的预测。文中以新浪微博实际数据为例，验证了基于异构信息网络分析的预测模型用于社交网络活动预测的可行性和准确性。更多还原

【Abstract】 In information network, a new heterogeneous trend comes true with thedevelopment of information diversification and complex relationships betweeninformation objects. Faced on the heterogeneous network structure, how to find theuseful knowledge and to improve the utilization of information based on networkanalysis is one of most urgent problems. In the heterogeneous information world, data,services, and human activities are the key engagement elements consisting of the usageof information. Relational database model can provide structured storage andmanagement format for the mass of information data; Web services as the representativefunction package and development technology can construct open, loosely coupledinformation platform; Social network provides the open platform for informationsharing and dissemination. With the diversification of data categories, frequent access toweb services and the rapid development of the social network, there is a growingdemand for information sorting and knowledge mining from the heterogeneous forms ofdata network, service network, and social network.In this paper, with the analysis of current information network and heterogeneousinformation network, faced on the problems of current researches, we mined therelationship between the heterogeneous objects in heterogeneous information network,studied of heterogeneous information network description model deeply, studied theheterogeneous information network description model based ranking functions whichconsider the linkage analysis of network structure. From the dimension of therelationship, the new heterogeneous information network analysis model basedclustering analysis, ranking and activity prediction methods are proposed. As the casestudy, we finished some basic researches and provide specific solutions for problems ofservices heterogeneous network, relational database tuple networks, social network.The details of research works in this paper include:①We analyzed the trends and heterogeneous characteristics of currentinformation network development, analyzed current situation and existing problems inclustering, personal query and social network prediction, Studied the format descriptionof heterogeneous information network description model.②Based on the heterogeneous information network description model, several network analysis based ranking functions are studied, considering the different forms ofnetwork connectivity and ranking rules; Through the comparison and analysis ofranking results, provide method support for network analysis based ranking.③Faced on the problems of property computing based clustering researches, formthe view of relationships we proposed a novel clustering algorithm based onheterogeneous information network analysis, and studied the basic idea and process ofthe proposed clustering algorithm; as special case of heterogeneous informationnetwork applications, in order to solve the problems of web services clustering, weproposed a new clustering algorithm based on service tags considered network structureand heterogeneous service network analysis, co-considering the clustering and rankingprocess of services, ranking model provides compute vectors for clustering process andfinish the ranking results in different service clusters. In order to evaluate theperformance and accuracy, we designed experiments with the true web services datasetfrom Titan.④In order to improve the personalized query support of in information searchesand queries, we proposed a personalized query method based on heterogeneousinformation network analysis which mines the possible ranking results considering thehidden categories. As the case study of specific network, we provided a new rankingalgorithm for personal top-k query in relational database which analyzes the foreignkeys linked tuple relations and schemas, study the extraction method of heterogeneoustuple network. Based on the relational tuples network structure, researched the rankingmodel and process of proposed algorithm. The ranking algorithm of relational databasecan classified as single category ranking and ranking consider multi-classes in whichthe latter ranking should consider the potential categories hidden behind data. Theexperiment analysis part chooses real databases from IMDB and German Credit toevaluate the proposed ranking algorithm on performance and accuracy.⑤Based on the means of heterogeneous information network analysis, combinedwith the social network activity forecast demand, we proposed social network activitiesprediction model based on heterogeneous information network analysis. Consideringthe analysis of network structure and messages propagation in social network, weproposed four different prediction models for message re-tweet and possible view times.Combining the proposed four prediction models with different weights based on accuracies of prediction, we constructed a composite prediction model. The experimentswith real Weibo data give an evaluation on available and accuracy of research.更多还原