The Research Of Community Detection In Heterogeneous Information Network And Its Application

Heterogeneous information network is a complex network with multiple types of nodes and links.These nodes and links contain rich semantic information,which brings more research opportunities and challenges to the researchers in data mining field.In recent years,researchers have made many achievements in similarity measurement,graph clustering,link prediction and recommendation in heterogeneous information networks.This thesis considers heterogeneous information network as the research target,and studies the relevant research mainly in the community detection and recommender system.For heterogeneous information network,there are three types of traditional community detection methods which are ranking-based methods,meta-path-based methods and multi-view learning methods.The first two methods are based on the probability graph model.The latter mainly uses the multi-view learning methods to solve the clustering problem in heterogeneous information network.And the recommender system based on heterogeneous information network can be regarded as the recommendation with multisource information fused,mainly to improve the performance of recommendation by fusing strategy and fused information.In contrast,this thesis starts with a new perspective(transforming heterogeneous information network into homogeneous information network).By means of the effective propagation of information through the meta-path,this thesis proposes a decomposition technique.In this way,original heterogeneous information network is decomposed into a series of homogeneous information networks with the benefit of zero-loss of information.At the same time,based on this decomposition strategy,this thesis proposes an algorithm called HomClus for heterogeneous information network and a recommended method CSR fusing user's and item's information.These three work constitute the core of this thesis which possesses the following major contributions:Firstly,this thesis proposes the meta-path-based decomposition strategy of heterogeneous information network.This method mainly makes full use of the property of the meta-path,that is,it reflect the nature of the different relations among entities.For entity with the target type,the relational weight matrices of entity with the target type under different meta-paths are obtained by simple matrix operation,that is,homogeneous information network.And the process is with zero loss of information for the target type.Therefore,the research problem on heterogeneous networks can be reduced to the research problem on the homogeneous network of entity with target type,which is easier to be solved.Secondly,based on the decomposition strategy of heterogeneous information network,a community detection algorithm,HomClus,is proposed.Firstly,the heterogeneous information network is transformed into a set of homogeneous information networks under the condition of the first contribution and then these homogeneous information networks are integrated into a unified network structure.Then,by exploiting nonnegative matrix decomposition,we project the nodes in network quickly into vectors,that is,the entire network is transformed to low-dimensional subspace.Finally,synchronizationbased clustering is applied to cluster the 'nodes' in the low-dimensional subspace and detects the potential community structure in the original network.Experiments show that the proposed method in this thesis has a great advantage compared with state-of-the-art algorithms.Besides,the algorithm is simple and concise,and the parameters are insensitive to the result.Beyond validity and practicability of the decomposition strategy of heterogeneous information network are also verified.Thirdly,the recommendation algorithm based on heterogeneous information network,CSR,is proposed.In this thesis,the heterogeneous information of the user and the heterogeneous information of the item are transformed into homogeneous information by using the decomposition strategy of the heterogeneous information network with the typical entity object,which are user and item in the recommendation system.Then combine the user information,the item information and the rating information together effectively by means of approximation with collective similarity regularization inspired by the popular recommender system method,latent factor model in collaborative filtering,and similarity regularization methods,and finally produce the high quality recommendation result.