The Research On Attributed Network Embedding Learning Based On The Dual Fusion Strategies And PPMI

Attributed network embedding(ANE)maps nodes in a network into a lowdimensional space while preserving the intrinsic essence of node attributes and network topology in the original network.In an attributed network,the attributes of node and the topology of network describe the same network from different aspects;the former portrays the node's individual profile from the micro perspective,and the latter illustrates relationships amongst the nodes from the macro perspective.Therefore,there exist the consistency and complementary information between node's attributes and network topology.The existing approaches can be divided into early fusion methods,synchronous fusion methods and late fusion methods according to the stages of fusion node attributes and network topology,i.e.,before,during or after the learning process.In fact,fusions at different stages have their own advantages and disadvantages.The single fusion strategy cannot take into account the consistency and complementarity of node attributes and network topology at the same time.Due to the heterogeneity of node attributes and network topology,directly fusing the two kinds of heterogeneous information is not conducive to fully mining the essential characteristics of the attribute network.In addition,most of the existing researches are only focus on homogeneous attribute network with the same node type and link type,ignoring the diversity of node types and link types in the network.To address the above challenges,the main work and innovations of this article are as follows:(1)This thesis proposes a homogenous attributed network embedding learning method named the ANEDF(attributed network embedding based on the dual fusion),which based on the double fusion strategies,i.e.,early fusion strategy and late fusion strategy.ANEDF consists of the early fusion component(EFC)and the late fusion component(LFC),where the EFC applied a shared Auto-Encoder to capture the latent complementarity between node attributes and network topology,and the LFC utilized two Auto-Encoder to extract the unique inherent essences from two types of heterogeneous sources.Auto-Encoder is the powerful model to learn the inherent essence from the input data.EFC and LFC are co-trained during the learning process,which promotes information interaction and mutual enhancement.(2)This thesis proposes a homogenous attributed network embedding method named the ANEDF(deep attributed network embedding based on the PPMI),which based on the PPMI(positive point-wise mutual information).DANEP first to construct the attribute graph based on similarity of nodes attributes,performs the random surfing on the attribute and topology graph obtain the attribute and topology PCO matrixes,and then calculates the PPMI on the attribute and topology graphs.Learning the low-dimensional representations from the cascaded PPMI of attribute and topology by the shared AutoEncoder.Converting attribute features to attribute graph can clearly describe the nonlinear manifolds structure of node attributes;the unified graph representations of attributes and topology help to integrate the complementary relationship between the two types of heterogeneous information;PPMI representations can capture the high-order proximity information and potentially complex nonlinear relationships of attributes and topology.Besides,DANEP designed a graph regularization with the local pair constraints based on attribute and topology graphs to enhance the consistency of local features in the process of learning low-dimensional representations.(3)This thesis extends the homogeneous attributed network embedding model DANEP to the heterogeneous attributed network that containing the multiple types of nodes and multiple types of link relationships,and proposes the PPMI-based heterogeneous attributed network embedding model named HANEP(heterogeneous attributed network embedding based on the PPMI).HANEP first to construct the attribute graph and extracts the complex topology information of the heterogeneous attribute network based on different meta-paths,obtains the PCO matrixes on attribute and multiple topology graphs extracted from the different meta-paths through random surfing,and then calculates the PPMI matrixes of attribute and multiple topology graphs.Learning the features essence of the multiple types of nodes node and the rich semantic information carried by the heterogeneous links between the multiple types of node by the multiple Auto-Encoders.Finally,cascading the low-dimensional representations from multiple Auto-Encoders as the embedding result of the heterogeneous attribute network that containing multiple types of nodes and multiple types of links.HANEP extends DANEP's advantages that attribute graph can clearly describe the non-linear manifolds structure of node attributes,PPMI representation can capture the high-order proximity information and potentially complex nonlinear relationships of nodes attributes and network topology,and the local pair constraints based on attribute graphs and topological graphs from different meta-paths to enhance the consistency of local features.(4)This thesis verifies the performance of homogeneous attributed network embedding models ANEDF and DANEP on the multiple real attributed networks by node clustering,node classification,link prediction,visualization task and parameter sensitivity analysis;and also verifies the performance of heterogeneous attributed network embedding model HANEP on two heterogeneous attributed networks by node clustering,visualization task,ablation experiment and parameter sensitivity analysis.Experimental results show that the performance of the proposed methods are better than that of the baseline algorithms.