Research And Implementation Of Graph Representation Learning Algorithm Based On AutoML

Graph data is ubiquitous in the real world,and we can use graph data to model complex relationships between different entities and entities,ranging from small molecules in proteins and particles in physical simulations to large power grids across the country And global airlines,it can be seen that efficient graph data representation is an important research direction in academia and industry.With the development of machine learning,machine learning on graphs,that is,graph representation learning,can efficiently mine the dependencies between nodes,capture information at the data structure level of graphs,and map each node in the graph from high-dimensional space to In the low-dimensional space,dense vectors are formed,so that the learned graph representation can be applied to downstream tasks.Although graph representation learning algorithms emerge in endlessly and all achieve good results in downstream tasks,existing research heavily relies on manual hyperparameter or architecture design to achieve the best performance.If we insist on using manual tuning when designing the best algorithm for the target task,it will lead to a lot of labor and material costs when there are a large number of models for various graph tasks.In recent years,automated machine learning(AutoML)has been extensively studied to reduce the human effort to develop and deploy machine learning models.The complete AutoML pipeline can automate every step of machine learning.Researchers only need to focus on the input and tasks to get the best machine learning model.Automatic machine learning on graphs,which combines the advantages of AutoML and graph representation learning,is gaining attention from the research community.This topic applies the idea of AutoML to graph representation learning,and designs an automatic graph representation learning algorithm.In hyperparameter optimization and neural network architecture search,it can target graph data in different fields and scales,and get good results in downstream tasks.performance.The main work of this project has the following three points:(1)A graph neural network architecture search algorithm based on a neural network predictor is proposed,which defines a search space that can generate topological structures,and initializes the node features and edge features of the architecture.In the search strategy,an attention mechanism-based neural network predictor considering different edge type information is proposed,and the architecture is efficiently embedded.Comparative experiments were carried out on the downstream tasks of node classification and link prediction,and ablation experiments were carried out through different predictors and architectural features to verify the effectiveness of the algorithm.(2)A graph neural network hyperparameter optimization algorithm based on subgraph construction is proposed,which uses the method of generating subgraphs to generate a graph summary that retains structural information for the original graph data,and expands the neural network predictor,based on the graph summary.Hyperparameter optimization process.The algorithm is compared and tested on different graph data sets,which proves the effectiveness of the generated architecture and the efficiency of the search time.(3)Design and implement the automatic graph representation learning prototype system,provide user management,experiment management,and algorithm functions,and provide users with a visualized,highly customized,one-stop automatic graph representation learning algorithm application platform.