Research On Robust Embedding Method Of Attribute Graph

In the real world,people’s daily life is producing a lot of data all the time.With the rapid development of Internet technology,it is possible to collect and mine these massive data.Among these massive data,there are quite a number of data that are related to each other.These data form massive attribute graphs(networks).Effective data mining on these attribute graphs can bring huge benefits to real life.Attribute graph embedding technology,as an efficient and mainstream graph mining technology,has been widely concerned in academia and industry in recent years,and has derived a large number of excellent attribute graph embedding methods and applications based on attribute graph embedding technology,which has brought great convenience to people’s daily life.However,most of the existing attribute graph embedding methods are based on the assumption that the graph data will not be contaminated.However,in real life,graph data is vulnerable to network attacks or anomaly node implantation,which will lead to a significant decline in the quality of node representation generated by attribute graph embedding methods.Therefore,this paper studies the robust attribute graph embedding technology,aiming at generating robust node embedding when the graph embedding model is polluted by the network.The specific work and innovation are as follows:1)Firstly,the reason why the existing methods are not robust is analyzed,because most of the existing methods focus on the local structure of the network,that is,reconstructing the connection properties between point pairs in the network.Therefore,when the network is attacked and there are attack edges(fake edges),the fake edges will seriously mislead the model.Therefore,this paper firstly proves that node attributes can distinguish between normal edges and fake edges,and proposes AGAT model.The AGAT graph embedding model incorporates node attributes in the attribute encoding and model training stages to ”filter” fake edges,so as to avoid the influence of fake edges on the graph embedding model and achieve the effect of generating robust embedding.Experiments and analysis on four real data sets have confirmed that the AGAT graph embedding model can produce high-quality and robust node embedding when the graph data has node attributes.However,when the graph data does not have node attributes,the AGAT model is difficult to use attribute information,and the effect is reduced.2)In order to solve the problem of poor effect of AGAT model in the absence of node attributes,an attribute graph embedding model An ECI based on preserving network community structure is proposed.Community structure is a more stable network structure than local structure,which is not easy to be destroyed when suffering from network pollution.Therefore,the node embedding model based on community structure guidance will be more robust.In the research,the traditional modularity function is analyzed firstly,and a new modularity function is proposed,which is suitable for more general scenarios.After that,the newly proposed modularity function is combined with the graph neural network as a loss,and the overall architecture of the An ECI model is proposed in combination with another loss.After that,experiments and analysis were carried out on four real data sets.The experimental results verified that An ECI can not only generate effective and robust node embedding,but also play an active role in each part,with high efficiency.