Traditional recommendation algorithms often face challenges such as data sparsity or cold start issues and may not effectively explore the correlation between rating information and auxiliary data. Convolutional Graph Neural Networks are broadly divided into two mainstream categories: spectral-domain-based convolution methods and spatial-domain-based convolution methods. In 2013, Bruna et al. published the first study on graph neural networks based on spectral-domain convolution. This study defined a specific way of graph convolution based on spectral graph theory. Since then, spectral-domain-based convolutional graph neural networks have received increasing attention . A series of spectral methods based on graph Fourier transform have been proposed. In these methods, spectral convolutional neural networks assume that the filters in the Fourier transform are a set of trainable parameters, laying the theoretical foundation for subsequent research. Following this trend, Kipf et al. proposed the groundbreaking Graph Convolutional Network (GCN), which efficiently approximates neighborhood convolution and aggregation by using Chebyshev polynomials to achieve a first-order approximation. The aim of this paper is to address the cold start problem using graph neural network technology while ensuring that the model retains the ability to leverage node content features. Additionally, the goal is to enhance the recommendation performance of the model on this basis, aiming to achieve or surpass the effectiveness of existing methods.