Abstract: We present a new geometry compression method for animations, which is based on the clustered principal component analysis (CPCA). Instead of analyzing the set of vertices for each frame, our method analyzes the set of paths for all vertices for a certain animation length. Thus, using a data-driven approach, it can identify mesh parts, that are "coherent" over time. This usually leads to a very efficient and robust segmentation of the mesh into meaningful clusters, e.g. the wings of a chicken. These parts are then compressed separately using standard principal component analysis (PCA). Each of this clusters can be compressed more efficiently with lesser PCA components compared to previous approaches. Results show, that the new method outperforms other compression schemes like pure PCA based compression or combinations with linear prediction coding, while maintaining a better reconstruction error. This is true, even if the components and weights are quantized before transmission. The reconstruction process is very simple and can be performed directly on the GPU.