Bipartite Network Embedding Based On Constrained Random Walk And Mixed Skip-gram

Network representation learning is a method to map a network to a low dimensional dense space,so that similar nodes in the network are similar in the low dimensional space,which is also called network embedding.Due to the difference of networks types,each network representation learning method focuses on different aspects.For example,in homogeneous networks,most methods only need to analyze the topological structure of the network,while in heterogeneous networks,they also need to pay attention to the heterogeneity of the nodes in the network.Bipartite network,as a special heterogeneous network,is one of research directions.The representation learning methods of homogeneous network and heterogeneous network representation are not suitable for bipartite network.At present,most of the bipartite network representation learning methods are based on the topological structure,few people consider the semantic information in bipartite network,meanwhile,it is easy to ignore that there are not only explicit relations with obvious links,but also implicit relations with same type nodes without links.How to effectively fuse topological structure information and semantic information is the key to learning the latent representation of bipartite network.In this thesis,we propose Bipartite Network Embedding based on Constrained Random Walk and Mixed-Skipgram(CRWMS)by analyzing the topological structure and semantic information of bipartite network.Our thesis adopts the random walk strategy based on constrained meta path to preserve the explicit and implicit relations between different or same type nodes.Then,the explicit path is transformed into two implicit paths with same type nodes by using separation operation.With these steps,we can not only get the topological structure information of bipartite network,but also the semantics information.As the training model of these random walk paths,the mixed skip-gram model breaks through the limitation that traditional skip-gram can only train paths one by one,and realizes the goal of joint training of implicit and explicit relations walk paths.At last,we obtain the effective representation vectors of bipartite network.Finally,the experiments of link classification,link clustering and recommendation are carried out on three real data sets,and compared with nine comparison methods.The results show that the network representation learned by CRWMS is more effective.In order to solve the problem of semantic confusion in walking paths,we further propose a method of bipartite network embedding with semantic enhancement(CRWMS++).The semantic information preserved by CRWMS is not clear enough,and it is easy to cause semantic confusion in some implicit paths.In order to solve this problem,CRWMS++ not only saves the topological structure and part of semantic information of bipartite network,but also preserves the rating value when generating the walk paths to enhance the semantic of paths,and both of them as the inputs of the mixed skip-gram with semantic enhancement model.In the process of training,a semantic parameter is defined to strengthen the semantic information,which is better integrated with the structural information.Compared with CRWMS,the bipartite network representation learning method with semantic enhancement is more superior in capturing rich semantics and fusing structure and semantics.