The Research On Linking Prediction Based On Subspace Clustering

The booming development of the Internet leads to the explosive growth of big data,resulting in large data scale and low value density,which has created a huge demand for effective and reliable data mining technology.Link prediction is to explore the possibility of the existence of relationships between entities,which has become the core task of data mining and has attracted extensive attention.Link prediction research advances the understanding of network evolution.However,the existing link prediction models are difficult to effectively model complex network structures.Due to the sparseness of the network,the numerical value density is low,and the noise interferes with the network structure greatly,resulting in low prediction accuracy.In particular,heterogeneous networks containing rich semantic information cannot make full use of diverse structural types,resulting in the absence or omission of key information.Based on the above reasons,thesis proposes a research on linking prediction based on subspace clustering.The specific contents of research are as follows:（1）For homogeneous graphs,the link prediction method based on subspace clustering is proposed.Firstly,a subspace clustering method based on low-pass filtering is proposed.There are high-frequency noise signals in the homogeneous graph network,which could reduce the accuracy of clustering.The low-pass filtering model can effectively filter the high-frequency noise in network,leaving the low-frequency real characteristic signal.At the same time,considering the influence of high-order neighbors on nodes,the learned network structure is helpful for downstream clustering tasks and link prediction tasks.Second,link prediction is performed based on the clustering information provided by subspace clustering and the common neighbors of nodes.Finally,the performance is improved compared to traditional local similarity methods based on common neighbors.（2）For heterogeneous graphs,the link prediction method based on subspace clustering is proposed.In order to take advantage of the rich semantic information contained in the heterogeneous information network,a meta-path-based similarity module is designed first.By learning the weights of different kinds of meta-paths,a self-expressive similarity matrix is constructed to complete the clustering task.Secondly,remove the heterogeneity in heterogeneous graphs.In other words,the heterogeneous graph is degenerated into a homogeneous graph.And the link prediction task is carried out by combining common neighbors and clustering coefficients,realizing the prediction in the heterogeneous network.Finally,the performance is improved compared to typical heterogeneous graph embedding methods.（3）Develop an academic network-based scholar cooperation link prediction system.The algorithm application has been implemented,and combined with the actual scenario requirements,a link prediction method based on subspace clustering has been designed and developed to develop a recommendation system for scholar cooperation link prediction under the academic network scenario,which realizes functions such as academic information query,scholar cooperation relationship prediction,and academic network visualization.In the future,the academic network scholar cooperation system can also provide effective ideas for the realization of scholar portraits and scholar association analysis.