Abstract: In order to enhance the interaction of points in DGCNN local graphs, the paper designs a new transition down layer to replace the EdgeConv module in DGCNN to accomplish the classification task. This paper used tangent vectors in the transition down layer to increase the geometric feature representation of points, grouped and sampled them by FPS and KNN algorithms, and inputted the sampled and grouped point clouds into the geometric affine modules GAM and MLP to complete the local feature extraction. This paper designed a shared edge function based on density-dependent statistics and a local grouping method based on K-Farthest algorithm. Classification experiments were conducted on ModelNet40, ScanObjectNN and ShapeNet datasets, and the proposed method achieved better results in all comparisons compared with DGCNN results.