| Community Detection is the foundation of data mining.Through community detection,you can understand the role of nodes in the network.At the same time,it also has a great guiding role in the mining,recommendation and prediction of the network's future node information.Most applicable environments of the existing community detection algorithms are homogeneous information networks,because the nodes of homogeneous information networks have a single node relationship,which can be intuitively described and easy to understand.However,with the development of networks,most existing information networks contain multiple types of nodes and link relationships.The link relationships between different nodes may represent different semantic information,and the same link relationship may have different expressions.Meaning,we call this type of information network a heterogeneous information network.There are some difficulties and challenges in selecting existing algorithms based on the homogeneous network community detection application in the existing information networks.Heterogeneous information networks are more comprehensive and accurate for node information expression,but they will face more difficulties and challenges than community discovery with homogeneous network structures.Based on the existing research results,this paper proposes a community discovery algorithm framework HCD_Clus(Heterogeneous Community Detection Cluster)suitable for existing heterogeneous information networks.The algorithm framework HCD_Clus mainly includes two parts: Heterogeneous Continuous Bag of Words-Similarity Measure Vector Autotropism(HCBOW-SMVA)similarity measurement algorithm that fuses multiple meta paths and a community based on seed node cluster Discover the clustering algorithm NS-Clus(New Similarity Clus).The HCBOW-SMVA algorithm is a similarity measurement algorithm that can be applied to heterogeneous information networks.The algorithm improves CBOW.It takes meta-path information as an input instance,obtains node vectors and meta-path weights,and then according to the node's own vector attributes.The similarity between nodes is obtained,and multiple meta paths are fused to obtain the final similarity measurement result.The NS-Clus algorithm first selects seed nodes based on node importance and second-order neighbors,and uses the module degree increment and the similarity between nodes based on the HCBOW-SMVA algorithm to perform community initialization of seed nodes.The clustering of non-seed nodes gives the final number of communities;secondly,the possibility of nodes belonging to the community,that is,the concept of membership,is used to add the remaining non-seed nodes to the community;finally,through community reconstruction,using improved labels Pass the idea of the algorithm to optimize the community,and at the same time find possible overlapping nodes in the community.The experimental environment of this paper is mainly the academic network DBLP and ACM datasets.These two datasets are typical heterogeneous network datasets.During the similarity index verification stage,AUC,precision,and NMI are used as evaluation criteria.The heterogeneous information network similarity index(HCBOW-SMVA)proposed in this paper is compared with traditional similarity index to verify its effectiveness.The experimental results show that the algorithm queries Top-k,verifies the accuracy of similarity,and verifies the validity of clustering results.All aspects performed well,indicating that the algorithm in this paper is feasible,effective and accurate for similarity measurement of heterogeneous information networks.The clustering algorithm NS-Clus is compared with some classic algorithms in community discovery.The mutual information NMI and the modularity Q are used for metric evaluation.The experimental results show that the NS-Clus algorithm proposed in this paper has higher accuracy and a more stable community structure,both in terms of clustering accuracy and the strength of the community structure.The performance is better,at the same time,the algorithm can also effectively find the overlapping nodes hidden in the community. |
