The Research On Relation Mining Algorithms Oriented To Social Network

Social network analysis has become a hot topic in data mining research.With the deepening of the research,the concept of social networks have been gradually expanded.At the same time,a large number of social networking platforms are developed,such as Facebook and RenRen that are based on a friend relationship,Twitter and Sina Weibo that aim to disseminating information,user-interest-based Flick and watercress,also a scholar DBLP based on cooperation and so on.Users produce a lot of entities and relations in the social network platforms every day.We can in depth understand of community structure of social networks,information dissemination,public opinion and swarm behaviors by mining the potential relationships between the entities.Therefore,the relationship mining between social networks have become the basis for related study.The relationship mining in traditional social networks focused on the relationship between humans.However,entities in real social networks not only refer to humans,but also many other entities.For example,the research for knowledge graph research mainly focuses on concept entities,information recommendation pays more attention to the item,location.The relationship in social networks include two categories: homogeneous relationship and heterogeneous relationship.Two types of relationships are synchronously exist and interrelated.Thus,entities and relationships are kernel of social network mining algorithms.From the perspective of data mining,the analysis of the complex relationships between entities in social networks is helpful to understand the socal network from the network structure,content and behavior.Our paper,focusing on the perspective of relationships mining,mainly conduct the research from the following aspects:First,a large number of homogeneous relations exist in social networks,such as the community features of users,the similarity of messages and the convergence of user behavior.Because homogeneity in social networks is difficult to be analyzed at a single scale analysis,a multi-scale analysis framework based on the diffusion wavelets is put forward.In the unified framework,multi-scale analysis is done for community structure of social networks,the topic,the user behavior.Secondly,a correlation matrix decomposition algorithm HSNMF-CM based on sparse non-negative matrix is propsed for solving large-scale heterogeneous relational data sparse and non-equilibrium problem.Select heterogeneous relations algorithm for a class corresponding to the smaller entities associated correlation matrix constructed,not only reduces the sparse matrix,but also improves the processing efficiency of the algorithm.In the block coordinate descent framework,HSNMF-CM Algorithm solves matrix factorization fast with efficient projection algorithm based on sparse constraints.For high-oder heterogeneous relations,HSNMF-CM Algorithm process complex heterogeneous relation data with clustering indication matrix fusion method.Thirdly,in oder to detect burst topic in large-scale microblogging message stream,the co-clustering algorithm is proposed to mine relations among users,message and entities based on dynamic window.Form the perspective of the entity influential,burst topic is defined with consideration to the dynamic particularity of topic.In the algorithm,the Chinese characters is used as entities.Then,using burst Chinese character build words and meaningful string.Finally,we can detect burst topics induced by new words and colloquial word.Finally,the anomaly detection algorithm based on heterogeneous relation matrix tri-factorization is proposed to solve the anomaly detection problem in microblogging.The similarity and dissimilarity homogeneous relations are measured quantitatively on aspects of content,interaction and user attribute.Then homogeneous relation is embedded into heterogeneous relation matrix with distance metric learning.The algorithm get anomaly detection result of users and messages through non-negative matrix tri-factorization on heterogeneous relation matrix.The algorithm achieves individual and group anomaly detection in microblogging.