In this work, we propose a new no-reference point cloud quality assessment (NR-PCQA) method leveraging a Graph Attention Network (GAT) to predict the visual quality of 3D point clouds. Unlike existing approaches, the proposed method does not require reference models, making it highly practical for real-world applications such as virtual reality, autonomous driving, and 3D reconstruction. The approach’s core involves extracting saliency maps from distorted point clouds, followed by clustering using advanced segmentation techniques. A weighted graph is then constructed, where nodes represent clusters and edge weights are determined based on the similarity between cluster centers using pairwise Euclidean distances. This graph structure is input into a GAT, which learns to predict the overall quality score by focusing on the most relevant relationships within the point cloud. We perform extensive comparative analysis on two standard subjective databases, SJTU-PCQA and ICIP2020. Results demonstrate significant improvements over state-of-the-art methods. Specifically, our model achieves a 10–15% improvement in prediction accuracy (PLCC) and monotonicity (SRCC) compared to existing NR-PCQA methods. These results highlight the effectiveness of the GAT in capturing local and global quality degradations in point clouds, making our approach a robust solution for point cloud quality assessment across various domains.