A Research Of Community Detection In Attributed Networks

With the development of technology,there are increasingly more kinds of network data which can be collected.Apart from the topology information among nodes,their attribute information are gradually accessed,which is called attributed networks.However,traditional community detection approaches only make use of the former while ignore the latter.Therefore,how to integrally utilize both of them for community detection on attributed networks is an important research problem in recent years.For network data,the homophily assumption(i.e.,nodes with similar attributes tend to link to each other)between topology structure and node attribute can be seemed as a bridge to integrate them.Based on this assumption,both topology structure and attributes of nodes can supply complementary information for community detection.However,due to the reason that there exist large noises in node attributes,directly utilizing all of them to measure attribute similarities among nodes will somehow lead to inaccurate results.As a common method for data mining,feature selection can effectively avoid the negative influence of attribute noises.However,it is commonly observed that edges largely exist among nodes from the same community,which is a local characteristic in topology structure.This characteristic accordingly leads the homophily assumption to be local,which means that there are different attribute sets that affect the links among nodes in different communities.Since feature selection fails to reflect the local characteristic,this paper proposes to only utilize attributes which are relevant to the formation of communities for detecting.Overall,the main works and their contributions are summarized as follows:(1)The construction of similarity matrix plays an important role in spectral clustering for community detection.To alleviate the negative influence of attributes which are irrelevant to the formation of communities,this paper proposes a spectral clustering based method in attribute subspace.This method measures similarities in attribute subspace and searches subspace via utilizing the two kinds of similarity,which is inspired by the homophily assumption.By iteratively detecting communities by spectral clustering and searching relevant attributes,the quantity of communities as well as precision of subspace are gradually improved.(2)Since traditional stochastic block models only consider the topology structure,this paper newly proposes a model to further consider the influence of relevant attributes on the formation of edges.Meanwhile,this method also propose to fit the model a constructed latent graph to avoid the negative sampling strategy which is widely used to solve the probability model.Comparing to traditional stochastic block model,this method can better model the connection pattern in node level and effectively explore the attributes which are relevant to the formation of community.(3)Since tradition random walk is only based on the topology structure.This paper newly proposes a random walk method which utilizes information of both topology structure and node attributes.Specially,this method obtain the relevant attributes for each node by an aggregating operation,which are then used to construct a bipartite graph.By walking on this bipartite graph,different nodes will automatically present discriminative distribution of transition probability.Comparing to traditional random walk based methods,this method not only reserves the characteristic of low complexity but also improves the quantity of detected communities.