Research On Application Of Graph Model Based On Bayes And Meta-learning

In recent years,the introduction of deep learning into graph structure data has aroused extensive interest among researchers.The search for better representation learning of graph structure has become a research hotspot.Among them,graph neural networks(GNN)is applied to research fields such as social network analysis,citation network analysis,and recommendation systems.Although many excellent models have emerged in the field of graph neural networks,and have achieved good results in solving intensive graph structure data applications,such as link prediction,node classification,and relationship extraction.But the traditional research methods use fixed learning algorithms to solve the task from scratch and require a large amount of data to be trained to achieve the desired results,and they need to relearn from scratch when faced with similar new problems.Although meta learning combined with graph neural network can quickly adapt to new tasks through few-shot learning,which can alleviate the problem of user data sparsity to a certain extent,but it often lacks good quantization of uncertainty.This thesis re-characterizes meta-learning from the perspective of Bayesian inference,and proposes a fusion algorithm based on gradient-based meta-learning and variational Bayesian inference,which can effectively amortize the hierarchical variational inference between tasks,learning the prior distribution on the weights of the neural network to produce good task-specific approximate posterior results,so as to achieve robust meta-learning in different graph models.The fusion algorithm is combined with the estimated graph structure and embedded in regression and classification tasks to provide the accuracy of the model and the interpretable basis for the application of graph data.Aiming at the task of graph model embedding regression,this thesis proposes a new inference model based on probabilistic meta-learning algorithm and graph neural network(Meta IMC)to deal with new user recommendation and cold start problems in matrix completion problems.The model can quickly adapt to the prediction of new types of users to improve the confidence of the regression model.To prove the effectiveness of Meta IMC,we evaluate the Meta IMC model on three benchmark data sets.The proposed model has achieved surprisingly generalization performance on matrix completion tasks.Aiming at the task of graph model embedding classification,this thesis proposes a general framework(Meta Geo)for identifying the user's geographic location,learning the prior distribution of the geographic location task,in order to quickly adapt to the user prediction problem from a new location.It solves the two limitations of the model proposed by the previous researcher that the number of samples is small and the user prediction model for new regions is not easy to promote.Meta Geo improves geographic location prediction under traditional settings by integrating a large number of micro-tasks.Probabilistic reasoning is integrated into the graph structure generation to solve the two inherent problems of position uncertainty and task ambiguity in the training of a small number of samples.In general,research on the application of graph models based on Bayesian and metalearning,using Bayesian statistical method to systematically explore the network representation of graph structure and information aggregation learning mechanism,and using variational inference and deep generative model to explore the learning principle of graph structure and model generalization performance.Thereby,uncertainty learning and model interpretability evaluation are further carried out,and a more efficient and understandable graph structure data prediction and classification model is established.