To address the limitations of point cloud segmentation methods based on the encoder-decoder architecture in terms of utilizing geometric details and contextual information, a point cloud semantic segmentation method based on graph pooling optimization is proposed to investigate the issue of feature loss. By analyzing the complementary relationship between geometric details and deep semantic features in the raw point cloud, a multi-level feature representation system is constructed. A technical solution combining the Edge-Guided Graph Pooling (EGP) module and the Edge-Guided Graph Unpooling (EGU) module is adopted. The EGP module explicitly retains edge structural features through a geometric constraint mechanism, while the EGU module guides feature reconstruction using the retained edge features, forming a closed-loop optimization system. Experimental tests are conducted on the S3DIS benchmark dataset, and the results show that the proposed method achieves a mean class accuracy (mAcc) of 73.8% on the Area-5 test set, which is a significant improvement over existing methods. Ablation experiments validate the effectiveness of the EGP and EGU modules. The method is capable of meeting the demand for fine geometric feature preservation in 3D scene understanding and demonstrates its potential for application in point cloud semantic segmentation.