Research And Application Of Multi-task-oriented Network Representation Learning Method

Networks can effectively describe complex systems in nature,where entities are represented by vertices and interactions between entities are represented by edges.In the real world,most networks typically contain multiple types of objects and the various links formed by their interactions,leading to the continuous increase in network size and the increasingly complex structure of the network.This type of network is also known as the heterogeneous information network.Many real-world tasks,such as citation network,social network,biological network and traffic network analysis,rely on graph data mining technology.Therefore,network representation learning has always been an important topic in network science,whose purpose is to design algorithms to identify potentially interesting graphic patterns,and to calculate,analyze and mine information networks in various scenes and tasks in the digital era.The basic idea of network representation learning is to learn the potential low dimensional representations of network nodes while preserving network topology,node content,and other information.This paper focuses on multi-task oriented network representation learning,and the specific research content is as follows:Firstly,this paper proposes an adaptive dual channel graph convolutional framework TADGCN for multi task learning.This model solves the problems of existing network representation learning algorithms based on two-stage or non-task-oriented and graph convolution in the fusion of node attributes and topological structures.TAD-GCN model not only considers task oriented,but also completes multiple tasks simultaneously by jointly optimizing multiple loss function.The information of nodes is obtained in topological space and attribute space respectively through the improved specific and public map volume layers.Correspondingly,consistency and difference constraints were designed for two types of graph convolutional layers.This overcomes the problem of graph convolution in the fusion of node attributes and topological structures.Numerous experiments on multiple real datasets have shown that the TAD-GCN model performs better than the comparison algorithm in multiple tasks such as node classification,node clustering,link prediction,and node visualization.Then,this paper proposes an adversarial heterogeneous network representation learning model TAM-AE for multi task learning.The existing learning models for heterogeneous network representation typically handle multiple links in the network through meta paths.However,this requires experts to define the meta paths of each dataset in advance,and algorithms based on meta path learning cannot fully capture information between nodes.To address these issues,the TAM-AE model first converts heterogeneous information networks into multi view learning from multiple network views,uses graph convolutional autoencoders to capture the consistency and complementarity information of each view,then uses adversarial training modules to force the potential representation matching prior distribution of each view,and finally introduces attention mechanisms to learn the associations between multiple views.The TAM-AE model is also oriented to multiple tasks,jointly optimizing the loss function of each part,and completing multiple tasks at the same time.The experimental results of multiple tasks such as node clustering,link prediction,and view reconstruction on multiple benchmark datasets show that the TAM-AE model outperforms the benchmark algorithm in terms of performance.