Path Laplacian Encodings for Directed Graphs
PathLapPE is a spectral positional encoding meant to keep directed graph structure intact instead of flattening it away.
The paper introduces node- and edge-level features derived from the path Laplacian on directed graphs. The authors say it captures directional higher-order structure and can be added to standard graph learning architectures. In benchmark tests, it improves node- and graph-level performance across several architectures, with the strongest results when paired with direction-aware message passing. Unlike magnetic Laplacian encodings, it does not require extra tuning of directionality hyperparameters.
ArXiv · AI/CL/LG's note
The paper introduces node- and edge-level features derived from the path Laplacian on directed graphs. The authors say it captures directional higher-order structure and can be added to standard graph learning architectures. In benchmark tests, it improves node- and graph-level performance across several architectures, with the strongest results when paired with direction-aware message passing. Unlike magnetic Laplacian encodings, it does not require extra tuning of directionality hyperparameters.
ArXiv · AI/CL/LG's note
score 4