| Network representation learning is a method to represent nodes with low-dimensional dense vectors to facilitate tasks such as community discovery and link prediction.With the development of deep learning,it can obtain more accurate vectors than traditional methods.Current research firstly conducted on static networks with nodes and edges that do not change with time,including homogeneous networks with the same node and edge types,heterogeneous networks with different types of nodes or multiple types of edges;Secondly conducted on dynamic networks and their research mainly focuses on dynamic isomorphic networks.However,most networks in reality belong to dynamic heterogeneous networks,which have both dynamic and heterogeneous characteristics.The existing dynamic heterogeneous network representation learning methods focus on the changes of the network as a whole at different times,but do not consider the changes in the influence of nodes during the evolution of the network,and the learning of heterogeneous semantics is relatively simple or the application scenarios are limited,which will lead to inaccuracy of the network representation vector.Aiming at the problems of existing algorithms,this paper studies dynamic heterogeneous networks.First,the concept and algorithm of node activity is proposed to measure the change of node influence in the network over time.Then,multi-subgraph semantic fusion is proposed.Algorithms are used to learn and fuse various semantic information contained in heterogeneous networks.The work of the thesis is mainly divided into the following four parts:(1)A dynamic heterogeneous network representation learning model NAMSF-DyHenet based on node activity and multi-semantic fusion is proposed.A dynamic subgraph set is obtained by performing various metapath walks on a dynamic heterogeneous network,and various semantic relationships are modeled into the subgraph set.In each subgraph,node activity and topological structure information are used for topology-level embedding,and then the semantic information of the subgraph set at each moment is fused.Finally,the dynamic evolution of the network is further learned through the recurrent neural network combined with the activity information.Through triple information capture of topology structure,logical semantics and temporal evolution,dynamic heterogeneous networks are comprehensively learned and accurately represented.(2)A dynamic network embedding algorithm based on node activity is proposed to learn dynamic network evolution patterns.Aiming at the change of node influence in dynamic network,paper proposes the concept of edge activity and node activity,updates the edge activity in the dynamic network evolution,calculates the node activity,and combines the recurrent neural network to learn the dynamic network evolution model.(3)A multi-subgraph semantic fusion algorithm is proposed to mine and fuse various semantic information in heterogeneous networks.The heterogeneous network is abstracted into a set of multiple semantic subgraphs through multiple meta-path walks,and the multiple semantic subgraphs are compressed using a shared parameter matrix.Semantic learning is then used to restore fully connected layers,and feature crossover is performed during the restoration process.The importance of different semantic relationships is learned through the process of compression and restoration,and semantic fusion is performed according to the importance.(4)Verify the effectiveness of the NAMSF-DyHenet model and the two algorithms.The paper uses three data sets of Aminer,IMDB and Yelp to conduct experiments and analysis on the model and algorithm.Compared with the other six comparison models selected in this paper,the results are 1%-8%and 0.4%-The accuracy rate is improved by about 12%,1%-22%,and ablation experiments are designed to verify the effectiveness of the two algorithms proposed in the paper. |
