Research And System Implementation Of Graph Neural Network Architecture Search

Graph is a common data structure in the real world,which abstracts entities as nodes and their interactions as edges.Various complex systems,such as social networks,protein networks,and traffic networks,can be modeled as graph data,making it convenient for researchers to analyze them using graph learning methods.With the rapid development of deep learning technology,graph neural networks,specially designed deep networks for graph data,have shown remarkable performance in various practical applications,such as social analysis,protein molecule prediction,and traffic flow prediction.Despite the great success of graph neural networks,a major challenge for researchers is that there is no single network architecture that can consistently outperform others on all tasks and datasets.It means that designing graph neural network models for new tasks and datasets requires a significant investment of computational resources and expert knowledge to search for optimal architectures,which can be time-consuming and labor-intensive,limiting the efficiency of graph neural networks in practical scenarios.To address this challenge,researchers have applied the idea of neural architecture search to the field of graph learning,aiming to enable machines to automatically design highperforming graph neural network models.However,existing graph neural architecture search works lack a comprehensive exploration of the highquality search space,the search strategies for complex heterogeneous graph scenarios,and the application of graph neural architecture search systems.To overcome these limitations,this paper focuses on graph neural network architecture search as the overall goal,conducting research from both algorithm design and system application perspectives.Furthermore,the former is further divided into two important parts:search space and search strategy,with algorithmic solutions or application systems designed accordingly.To explore the search space of graph neural network architectures,focusing on the application of recommendation systems,this paper first summarize the unified design framework of graph neural network based collaborative filtering and extract key search dimensions and common ranges of values to form a vanilla search space.Then,we conduct experiments to evaluate the vanilla search space,present the evaluation results and analysis,and use these results to prune the vanilla search space.Finally,we demonstrate the advanced experimental results achieved through architecture search in the pruned search space to validate its effectiveness.To explore the search strategy of graph neural network architectures,this paper focuses on heterogeneous graph neural networks.We investigate how to design architecture search strategies that effectively and adaptively utilize key semantic information in heterogeneous graphs.We introduce the node dependent semantic search solution,i.e.,NDS,and demonstrate its advanced experimental results on multiple datasets in two mainstream graph learning tasks,node classification and link prediction,and conduct result analysis.To explore the system application of graph neural network architecture search,this paper designs and implements a graph neural network architecture search system that integrates the aforementioned algorithmic achievements.It provides users with an interactive interface and a graphical display page for graph neural network architecture search,reducing the threshold for the use of graph neural network architecture search technology and promoting its practical application.