Following the introduction of graph neural networks, researchers have discovered the potential of using these networks to extract interconnections between users and items, which are often challenging for traditional recommendation algorithms to capture. The use of graph neural networks in constructing recommendation algorithms has gradually become a current trend in recommendation algorithm research. Currently, most recommendation algorithms based on graph neural networks extract user and item features using ID information, neglecting the implicit user and item features present in other data such as textual reviews. Alternatively, some algorithms incorporate auxiliary information, but this often complicates network training. To address this, our paper proposes a recommendation algorithm based on a graph neural network that integrates traditional deep learning, combining different user and item features. We introduce an attention mechanism to capture higher-order characteristics between users and items as part of the user and item feature representation. Simultaneously, we employ a traditional deep learning network to extract user and item features from textual data. The final user and item feature representation is obtained by merging these two feature representations, leading to improved recommendations. Experimental results demonstrate that this method effectively integrates high-order features and general characteristics of users and items, enhancing the representation of user and item features and consequently improving recommendation performance.