| Due to the huge increase of transmission speed in computer networks,data flow between entities in the real world is becoming more and more frequent,and the era of big data has come.As a common method to study the characteristics of interactions,organization,and other relationships between entities,graph modeling is of great importance in the fields of data mining and analysis.In addition,due to the powerful advantages of neural networks in feature encoding,many graph neural networks have been proposed to achieve tremendous success on graph downstream tasks,such as recommender systems and knowledge graphs.As the primary aspect of graph data analysis and processing,graph representation learning aims to encode node attributes and behavioral relationships to learn vector representations,and many practical application scenarios have demonstrated the importance of representation learning for graph downstream tasks.However,the current mainstream methods have some problems,such as the fixed propagation of graph convolutional networks cannot dynamically control local circulation,traditional aggregators are prone to over-smoothing when the depth of the model gradually increases,and the regularization reconstruction in encoder-decoder can prevent the model from learning differentiated node features,which may reduce model encoding ability and node representation adaptability to graph downstream tasks.Considering the limitations of the current work,this paper proposes a multi-level feature augmented deep graph representation learning model MFADGI(Multi-level Feature Augmented Deep Graph Infomax for representation learning)based on the theory of mutual information maximization,which can generate high-quality nodes in an unsupervised contrastive learning manner,and synthetically improve the performance of graph downstream tasks.First,a feature extractor composed of a multilayer feedforward neural network is designed,so that MFADGI can preserve the differentiated features in original node attributes,and initially enhance node semantic representation itself.Secondly,the node extraction representation exported by the extractor is fed into a deep graph neural aggregator containing attention layers,and through propagation and iterative update of node representations,the node aggregation representation exported by the aggregator can acquire more higher-order similar features in a wider range,further enhancing the semantic expression of the node representation for contextually relevant nodes.Furthermore,to solve the problem of over-smoothing caused by the deep propagation,this paper proposes a simple and efficient jump connection method,which allows nodes to acquire higher-order features while maintaining independence constraints to avoid excessive loss of their original feature information.Finally,MFADGI introduces the strategy of DGI to keep consistent encoding rules for all nodes and improve the generalizability of node representation for downstream tasks.Experiments demonstrate that MFADGI outperforms mainstream comparison baseline on both transductive and inductive node classification tasks in several benchmark graph datasets,and achieves optimal experimental results on link prediction tasks as well.Several performance evaluation experiments demonstrate the effectiveness of the MFADGI model proposed in this paper,which has significant advantages on graph representation learning tasks. |
