Research On The Learning Method Of Heterogeneous Network Representation Fusing Multiple Semantic Element Paths

A heterogeneous network is a kind of complex networks and contains multi types of nodes or edges.Its complex and diverse relations make it more vividly to model heterogeneous multisource information of the real-world.Therefore,the research of heterogeneous networks has become a hot topic in the academic area and industry.However,due to the complexity of heterogeneous networks,heterogeneous network embedding has become one of the most challenging essential research topics in the field of complex networks.Recently,many heterogeneous network embedding algorithms have been proposed based on the idea of ‘divide and conquer',in which a heterogeneous network is divided into several homogeneous sub-networks,and then the embeddings of each homogeneous sub-network are learned using homogeneous network embedding algorithms.Different embeddings of subnetworks are finally fused to obtain the final embeddings.This kind of heterogeneous network embedding algorithm has two challenges.One challenge is learning more accurate node embeddings of homogeneous sub-networks.Although significant progress has been made in the homogeneous network embedding,most of the proposed methods are designed for unsigned networks and not suitable for signed networks.In signed networks,there are not only positive links but also negative links,which are also very important in network structures.The other challenge is fusing multi-source information of heterogeneous networks to obtain more comprehensive embeddings.The existing heterogeneous network embedding algorithms are mainly fused the explicit meta-path-based semantic information.There is still much implicit information in the network,such as the implicit meta-paths caused by influence diffusion,the attribute information of nodes,and the attribute information of relations.How to fuse those mentioned above heterogeneous multi-source information to learn more accurate embeddings remains to be studied.The paper mainly focused on representation learning of signed networks and the fusion of heterogeneous multi-source information.In this paper,a multi semantic meta-paths fusion-based weighted signed heterogeneous network embedding model was built to learn polar relations,fuse influence diffusion meta-paths,and fuse multimodal information.The main contributions of this paper are summarized as follows.(1)It proposed a signed network spectral embedding method based on average commuting time.The paper focused on the key problem of signed network embedding on how to keep the negative relation of signed networks in low-dimensional embedding vector space and proposed an average commute time-based spectral embedding method for signed networks(abbreviated as CDSNE).It designed a random walk model in signed networks that a walker traveled along the positive edge with a higher probability and along the negative edge with a lower probability to increase commute distance of nodes linked by negative edges.The obtained nodes sequences preserved the first-order proximity of signed networks.Based on the relation of average commute time and Laplace spectral,the paper proposed an extended Laplace spectral and proved that the learned spectral embedding could maintain the average commute time of signed networks.Experimental results demonstrated that the obtained low-dimensional vector representations could accurately predict the sign of links.By setting a reasonable constricting factor,the average commute time of negative links can be widened,and community structure can be detected.(2)It proposed a signed network embedding method based on random walks guided by the second-order neighbor.The above-proposed spectral embedding method maintains the first-order proximity of the signed network and is not suitable for large-scale signed networks due to its high time-consuming.The paper proposed a signed network embedding method based on a random walk guided by the second-order neighbor(abbreviated as NRW-SNE).A new bias random walk procedure was designed to obtain nodes sequences.In the random walk procedure,a walker can walk along with the first-order neighbor or the second-order neighbor.The node sequences preserved the first-order proximity and second-order proximity of the signed network.The embeddings were learned by maximizing the occurrence probability node pairs in the sequences to improve the efficiency of the proposed embedding method.The experimental results showed that the learned embeddings are better than spectral embeddings in community detection and signed prediction tasks.(3)It proposed a new heterogeneous network embedding method based on influence diffusion meta-paths fusion.The existing heterogeneous network embedding algorithms lack fuse implicit meta-paths caused by influence diffusion.The paper proposed a heterogeneous network embedding method based on influence diffusion meta-path fusion(abbreviated as ID-HNE).It gave a formal definition of the influence diffusion meta-path and calculated the relation matrix of the meta-path generated by a k-step influence diffusion,and found that the fused relation matrix is equivalent to Katz similarity and utilized the Katz similarity to fuse the influence diffusion meta-paths into a single meta-path.Then it built a semi-supervised stack de-noising autoencoder framework to learn the embeddings of each meta-path.The embeddings were deeply fused to obtain the final embeddings.The experimental results on drug heterogeneous network and recommendation system verified the effectiveness of the learned dug embeddings.(4)It proposed a weighted signed heterogeneous network embedding method based on multi-source information deeply fusion.There are many types of heterogeneous information,such as node attribute and relation attribute in heterogeneous networks.The paper proposed a weighted signed heterogeneous network embedding method to learn the embeddings of users and items in the recommendation system based on multi-source information deeply fusion(abbreviated as WSHE).It defined weighted meta-path-based proximity to measure the polar relation of user preference.A heterogeneous network was divided into several homogeneous sub-networks based on the defined meta-paths.In every homogeneous sub-network,the embeddings of users or items were learned by maximizing the occurrence probability of node pairs in sequences that were obtained using the defined weighted random walk procedure.In the fusion stage,the method adopted the attention mechanism and pooling operation to fuse semantic meta-path information and attribute information and learn the final embeddings.The model was optimized by rating prediction tasks in the recommendation system.Experiments showed that the learned embeddings of users and items could effectively improve the recommendation performance.