Hyperbolic embedding distortion and Gromov delta-hyperbolicity fit

methods / Theoretical / Analytical / Geometric analysis / Hyperbolic embedding distortion and Gromov delta-hyperbolicity fit

Hyperbolic embedding distortion and Gromov delta-hyperbolicity fit

Techniqueadvanced

Measures whether a network's learned embedding (typically a hyperbolic GNN's Poincare-ball or Lorentz-model latent space) actually achieves a low-distortion, metric-faithful embedding of the graph it is trained on, via a contraction/expansion ratio between learned latent distance and graph shortest-path distance, cross-checked against each dataset's own exact Gromov delta-hyperbolicity.

Used in (1 observation)

Hyperbolic GNNs achieve low-distortion embeddings only on tree-like graphs under geometry-aligned tasksmeasured

structure: Curvature profile of the representation manifold · models: HGCN (Hyperbolic Graph Convolutional Network, Poincare ball, learnable curvature), HyboNet (fully hyperbolic GNN, Lorentz model) · paper: Hyperbolic Graph Neural Networks Under the Microscope: The Role of Geometry-Task Alignment