A Research On Graph Neural Network Based On Contrast Learning

Graph is composed of nodes and edges and describes the general connection between different entities.Graphs can be divided into homogenous graphs and heterogeneous graphs according to whether the types of nodes and edges are single or not.In order to mine different graph structure data,graph neural network has become the current mainstream means,which aims at continuous information passing and aggregation,to obtain high quality nodes representations.In order to alleviate the dependence of graph neural network on artificial labels,contrastive learning,as a typical technique of self-supervised learning,has become a research hotspot in the academic circle by learning the discriminative node representation through the spontaneous positive and negative case relationship of the comparison data itself.In order to further explore the nature of graph contrast learning and its practical application value,this paper reviews and improves the existing graph contrast learning framework from the aspects of theory,method and application,strengthens the deep integration with graph neural network,and realizes the practical industrial implementation of this technology.At the theoretical level,this paper explores the graph contrastive learning augmentation strategy based on spectral graph theory.First of all,the GAME rule is revealed through experiments,that is,if the difference between low-frequency part of the two input augmentations is less than that between the high-frequency information,they will improve the effect of contrastive learning.Such two augmentations are defined as optimal contrastive pair;Then,the correctness of GAME rule is proved experimentally and theoretically,and it is pointed out that contrastive learning can make the encoder pay more attention to the invariant information between two augments.Based on GAME rule,this paper designs an general plug-in model SpCo,which can be combined with the existing graph contrastive learning model to effectively improve the effect of the base models.Finally,SpCo is combined with other baseline models to verify the enhancement effect of SpCo on the datasets.At the method level,this paper studies the homogenous graph structure learning based on contrast learning.Firstly,this paper divides the graph structure learning model into two branches based on single view and multi-view,and focuses on the optimal graph structure in the multi-view based graph structure learning.Secondly,with information theory,this paper formally defines "minimal sufficient structure" as the optimal graph structure,which can take both effectiveness and robustness into account.Then,we prove how to obtain the minimal sufficient structure,and the multi-view graph structure learning model CoGSL is designed under the guidance of this theory.Finally,the node classification accuracy and robustness of the proposed CoGSL model under random attacks are verified on multiple datasets.At the method level,this paper also designs a heterogeneous graph neural network based on contrastive learning.This paper is the first to apply the contrastive learning technology to heterogeneous graphs,and designs a heterogeneous graph neural network model HeCo based on collaborative contrastive learning.First,the model depicts the metapath view and the network schema view.In order to increase the difficulty of contrastive learning,the view mask mechanism and two generations of difficult negative samples are designed.Secondly,according to the semantic information of heterogeneous graph,a collaborative optimization mechanism is designed to capture the commonness between the two views.Then,considering the view-specific information inside each view,this paper further proposed the HeCo++ model,capturing inter-view and viewspecific information simultaneously.Finally,the validity of the two models is proved by experiments,and the potential of contrastive learning in heterogeneous graphs is revealed.At the application level,this paper applies graph contrastive learning to the social recommendation problem for inactive users.Firstly,this paper analyzes the characteristics and shortcomings of current social recommendation models and points out that they do not consider the quality of raw social relations.Secondly,by observing the user behavior data in real industrial scenarios,it is found that inactive users generally have fewer social relations,and some of thm are inferior.At the same time,it is found that establishing relationships with active users can bring more positive gains;Based on the observation from industrial dataset,this paper designs the social recommendation model LSIR for inactive users,aiming to refine existing social relations by using graph structure learning technology,and proposes the mimic learning based on graph contrastive learning to strengthen the modeling of social relations for inactive users,so as to improve the recommendation performance on them.Finally,this paper has achieved a great improvement on both public and industrial datasets,which verifies the necessity of refining the structure of the raw social graph.