Network Representation Learning And Link Prediction Towards Attributed Network

With the explosive growth of information,analysis and mining of large-scale graph like social network and biological network have attracted increasingly much attention.Several problems such as graph partitioning,clustering,link prediction and community search etc.have developed as independent research fields.Network representation learn-ing,named as network embedding as well,aims at mapping nodes into continuous and low-dimensional vectors,and then traditional vector-based approaches such as clustering,classification can be utilized.Hence network representation learning has been regarded as the basis of many other graph mining work and has considerable research significance.However,the majority of existing graph embedding models solely consider the topolog-ical structure of network and ignore abundant content information of nodes.To improve the performance of traditional models,we proposed the attributed graph embedding model based on random walk procedure and word vector model.Besides nodes,our model can also learn mapping of attributes into low-dimension vectors.With vectors learned in the model,we further presented a fast link prediction algorithm.Main contributions of our work can be summarized as follows:·Attributed network embedding model We proposed attributed network repre-sentation models AttrCBOW and AttrSkipGram which focus on mapping nodes and attributes into low-dimensional vectors.Vectors we obtained are structure-aware and attribute-aware compared with original network.· Attributed network link prediction algorithm We posed a link prediction algo-rithm based on vectors.In order to improve the efficiency,BMH(Balanced Min-Hash)was designed to generate signature matrix and then LSH(Local SensitiveHash)will be used to reduce candidate pairs.· Network embedding model evaluation In the experiment of multi-label classifi-cation,we compared AttrCBOW and AttrSkipGram with traditional content-based and structure-based models.The results demonstrate that our models remarkably outperform baselines in accuracy,robustness and rate of convergence.· Link prediction algorithm evaluation Experiments illustrated the effectiveness of our link prediction algorithm.The combination of LSH and BMH can distinctly filter a large proportion of dissimilar node pairs.