Research And Application Of Weak-supervised Graph Representation Learning Algorithm

In the context of the Big Data era,which pursues the interconnectedness of everything,various kinds of network data have exploded,with social networks,mobile communication networks,logistics networks and other network data constantly emerging,making information networks an extremely common carrier and form of information in production and life.Graph representation learning can learn and analyse the intricate interactions in network elements to achieve more accurate causal and associative reasoning,so graph representation learning has great application value in node classification and clustering(social networks,biomedical),link prediction(commercial search,recommendation),etc.However,traditional graph representation learning algorithms rely on a large amount of reliable label information,ignoring the problem of sparse label information and the presence of noisy information in realistic scenario networks.To address this problem,this paper conducts research and implementation of the following:(1)In this paper,we propose a sparse label-based graph representation learning algorithm model ST-LPGCN,which uses a multi-stage selftraining approach to generate pseudo-labels,expand the label set and enhance the supervision information;adopts a graph convolutional neural network to learn local structure information and node attribute information,while using a label propagation algorithm to capture the global structure information of the graph,both of which generate pseudo-labels simultaneously;designs a pseudo-label comparison mechanism to reduce pseudo-label noise and improve the confidence level of the final pseudolabels,thus effectively improving the performance of the graph representation learning algorithm model.(2)We propose a label-noise-resistant graph representation learning algorithm model GCMT,which expands the positive sample views of the graph by data augmentation through techniques such as graph diffusion,edge perturbation,attribute masking,and subsampling;the average teacher model is used as the main framework,and the multiple views generated by the data augmentation module are divided as the input of the teacher network and the student network for training and learnling,and the output of the teacher network is The output of the teacher network is a "soft"pseudo-label,which is used as a supervisory signal for the student network;the output of the two networks will be compared and learned at the node level to improve the robustness of the node representation to mislabelled node information;at the same time,the deep clustering module clusters the feature vectors learned by the teacher network to generate a "hard" pseudolabel,which is also used as a supervisory signal for the student network."pseudo-labels,which are also used as supervisory signals for the student network.(3)A graph representation learning system is designed and developed,integrating the algorithms studied in this paper and other mainstream graph representation learning algorithms;at the same time,a variety of basic algorithm components are integrated to achieve data management,data analysis and processing,result visualization and other functions,and provide a friendly interactive interface to promote the development and application of graph representation learning algorithms.