| Graph is a very common and valuable data,which can be found everywhere in the era of big data.Graph neural network is artificial intelligence models developed for graph data.Self-supervised learning aims to reduce the use of labeled data by pretraining the model using a large amount of unlabeled data and fine-tuning it in downstream tasks using less labeled data.Incorporating self-supervised learning into graph neural networks can improve the performance of graph neural network with less use of labeled data,so it’s important in graph neural network research field.Existing graph self-supervised methods are mainly divided into contrastive and generative.Contrastive self-supervised methods rely on augmented data,but due to the specificity of the graph structure,existing methods use a lot of manual or skillful ways to construct graph augmented data,which increases the complexity and reduces the generality of the methods.Generative self-supervised methods mostly use autoencoder and other methods to reconstruct the input data,but generative models require a large amount of training data and demanding skillful design of the generation task.To solve the above problems,this thesis proposes generative graph self-supervised model based on multilayer graph neural network feature representation and contrastive graph self-supervised model based on mask mechanism,and conducts experiments on commonly used graph datasets.The main work of this thesis is as follows:1.Proposing the generative graph self-supervised model based on multilayer graph neural network feature representation.The architecture is autoencoder,which masks the graph nodes’ attributes randomly and utilizes encoder’s intermediate layers’ information.The decoder reconstructs the masked nodes’ attributes and compares the reconstructed nodes’ attributes with the original nodes’ attributes.This model allows the encoder to learn rich information of the graph structure as well as increase the robustness of the model.Experimental results on relevant graph datasets show that this model has advantages over other generative graph self-supervised models and outperforms some graph supervised models in terms of performance metrics.2.Proposing the contrastive graph self-supervised model based on mask mechanism.This model adds contrastive module on the previous model,and uses the mask mechanism to generate augmented data quickly and efficiently.This model integrates contrastive and generative self-supervised learning,enabling graph neural network to learn more comprehensive and rich graph structure information.Experimental results on relevant graph datasets show that this model has advantages over other popular contrastive graph self-supervised models and is competitive with some supervised models. |
