Research On Deep Neural Networks Based Classification And Representation Learning Of Heterogeneous Networks

In reality,most real systems usually contain a large number of interconnected,various types of components.Most research work models these complex systems as homogeneous networks,and ignores the complexities and semantic relationship between the components.Recently,more and more research work has realized that there are interconnected,complex,and multiple types of objects in real networks.Therefore,they have modeled these complex systems into heterogeneous networks(Heterogeneous networks,HNs).However,the heterogeneous network contains a large number of different types of objects and the relationships between objects.How to use the complex semantic relationships in the heterogeneous network and fuse various information in the heterogeneous network face a huge challenge.The challenges are as follows: 1)Facing with the growth of massive data and the mining of associations between different types of data,how to model these complex and different types of objects is an urgent problem.2)How to fuse different types of data in a large heterogeneous network to obtain effective information is still lacking effective solutions.3)Heterogeneous networks contain many different types of vertices and link relationships between vertices.How to accurately utilize these complex semantic relationships is still a hot and difficult issue in the research work of heterogeneous network.To solve the above challenges of heterogeneous network research,this paper first considers the representation learning of heterogeneous network nodes,then combines deep neural networks to process complex data of high-dimensional structures in heterogeneous networks,and finally validates our proposed model in semi-supervised learning tasks.The main research contents and innovative work of this paper are as follows:1.First,we extend the labeled objects in heterogeneous information networks based on some rules.Then,we use Meta graph-based Classication of Heterogeneous Information Networks(MCHIN)to measure the relatedness between the same or different-typed objects.Finally,we grouped the objects into pre-specied classes by using the ranking distribution of objects.The network structure employed by the ranking model can be adjusted so that each class become more distinguishable.We conduct comprehensive analysis on the proposed model in order to gain more insights from datasets.The results show the effectiveness of MCHIN in the classification task.2.To solve the noisy and sparse data,especially alleviate the “block” problem during the hierarchy construction processing in heterogeneous network,we put forward the idea that make use of both Stacked Denoising Auto Encoder with sparse factors(SDAEf)and the relax strategy(RSDAEf)to construct hierarchical structure,finally to classify the objects.Firstly,we design the SDAEf model to extract features of objects in heterogeneous networks,which can figure out the noisy data.Our approach retains the benefits of Stacked Denoising Auto Encoder(SDAE),and extends it with sparse factors to deal with noisy and sparse data.Then,to alleviate the “block” problem during the hierarchy construction,RSDAEf is proposed to construct hierarchical structures by the relax strategy.Finally,we conduct experiments on real datasets to evaluate our algorithm.The experimental results show that our methods are highly effective and achieve high classification precision in heterogeneous networks.3.One popular tool to study HNs is the meta-paths,which could better preserve relationship information between different types of entities in HNs in comparison to random walks.However,a meta path itself has limited semantics,since the length of meta-paths are no more than four in most studies.Another problem is that typical meta-path generation can produce many short paths or even isolated nodes which are hard for further process and learning.Therefore,to take full advantage of the meta paths and remedy the gap of different objects,we propose the weighted meta-graph to better capture graph relationship semantics.Next,we propose a modified version of Graph Convolutional Network(GCN)to further process the results from the meta-graph to learn HNs embeddings.The weighted meta graph makes up for the lack of semantics in GCNs,while the GCN mixes the network topological structure information of with the semantics.Besides,we modify the convolution operator to consider what we named as node “self-significance”-how much a node should consider its own feature when convolve with its neighbors through the iterations.Finally,extensive experiments are done to compare with a variety of recent models on four real-world data sets,with different train-test splits,to confirm the effectiveness of our approach in classification and link prediction.4.Aiming at the neglect of the role of the neighbors of different types of nodes in the heterogeneous network,and considering the fusion of target node neighbors,this study proposes to use the attention mechanism,graph convolutional neural network,and Bidirectional Encoder Representations from Transformers(BERT)model to extract features of heterogeneous networks.Firstly,to extract local neighbor features and learn the importance of different neighbors of the central node,graph convolutional neural network is used to fuse the attention mechanism.Targeted learning of different roles between different neighbors through attention mechanisms.Secondly,to learn the potential distribution of node features,a new adversarial regular model is proposed.The model can learn the potential distribution of vertex features by creating an error between the generated potential distribution of vertices and prior knowledge.Finally,in order to extract the features of the heterogeneous network and make up for the shortcomings that the attention mechanism cannot extract the data features over a long distance,the random walks of the graph are used to obtain a randomly generated vertex sequence as the input of the BERT that incorporates the graph convolutional neural network.Experiments on bioinformatics networks,social networks,bibliographic information networks,etc.,verify the robustness and universality of the method.