We are now entering the era of big data, transitioning from a previous era of information scarcity to one where we face significant information overload. In this context, the challenge of filtering out valuable information for individuals from the vast sea of internet data has become increasingly important. A graph is a crucial information organization structure consisting of nodes and edges. Content-based recommendation systems require extensive background knowledge about items and users. This knowledge can often be organized in the form of a knowledge graph. With the development of the internet, a considerable amount of knowledge has been structured into graphs, such as Wikipedia or knowledge trees and classification tree structures like e-commerce directories. This paper conducts research on recommender systems based on graph neural network (GNN) technology. Utilizing graph convolutional neural networks, a graph information extractor is constructed to merge information from the user-item bipartite graph's node neighbors and generate dense vector representations for the nodes. The model employs a multi-task learning approach by introducing the reconstruction tasks of auxiliary information graphs for users and items into the traditional collaborative filtering recommendation task through graph neural networks.