Research On Neural Network Based Graph Representation Learning

With the popularization of Internet devices and mobile terminals,the era of big data has profoundly changed people's life style.People use the Internet to carry out education,medical care,shopping,entertainment and other activities conveniently,accompanied by massive data.Among them,graph data widely exists in various scenarios,such as financial network,social network,medical graph,which are often characterized by large scale,complex structure,multi-domain information,etc.,resulting in unsatisfactory algorithm performance and time cost in the process of analysis,which seriously affects the overall performance of related tasks.In order to effectively and efficiently process graph data to overcome the above challenges,and apply it to different scenarios to support different downstream tasks and improve the availability of graph data,the main work of this thesis is as follows:(1)In view of the high dimension and sparsity of graph data,a graph representation learning method based on diffusion mechanism is proposed.Inspired by the diffusion phenomenon,this method utilizes diffusion system to characterize the information flow on graph,then to better model local and global properties of graph.In order to establish the relationship between feature attributes and structure attributes,a heuristic structure self-refinement algorithm is proposed to feed back the node representation results to the graph structure.In addition,the degree penalty is introduced to ensure that the scale-free feature of the real graph can be maintained in the representation space.Extensive comparison and visualization experiments demonstrate the effectiveness of proposed method.(2)In this thesis,a two-stage network,DAGDA,is proposed to solve the problem of class template distribution irregularity and graph representation learning of bipartite graph with heterogeneous nodes.In the first stage,the network extends the spectral domain graph convolutional network to aggregate the information of nodes on the homogeneous heterogeneous graph so as to output the discriminative class representation.In the second stage,a projection network based on auto-encoder is proposed to implement the classification under zero-shot situation.Meanwhile,the relation regular constraint is applied to the second stage network,and the distribution of samples in attribute space and representation space is aligned in a fine-grained manner.In order to evaluate the performance of proposed model,lots of experiments are carried out in this thesis.The results show that DAGDA is comparable with the state-of-the-art models in several benchmarks.(3)To solve the problem of sparse and cold start in recommender system,a Knowledge Interest Network(KIN)is proposed to introduce knowledge graph as auxiliary information.The network learns entity representations of knowledge graph and prediction tasks of recommendation system in a end-to-end framework.In order to learn entity representations,a spatial convolutional network is constructed to improve the efficiency of convolution operation on large-scale graph data.In addition,in order to achieve better graph convolution effectiveness,an entity translation based attention mechanism is proposed for effective information aggregation between users and items.The evaluation result of proposed model on the benchmarks is significantly better than that of the recent recommendation algorithm,which verifies its effectiveness.