Notes
🔍

Geodesic Distance

When data lives on a curved shape, straight-line distance can be misleading. Geodesic distance is the distance you get by traveling along the surface or structure of the data, rather than cutting straight through space.

What it means

In manifold learning, the key idea is that high-dimensional data points may lie on a lower-dimensional curved surface. The ordinary Euclidean distance between two points measures the direct shortcut through the ambient space, but that shortcut can ignore the true geometry of the data. Geodesic distance measures the shortest path constrained to the manifold itself. A common analogy is walking on the Earth: two cities are not connected by a straight line through the planet, but by the shortest route along its surface.

How it is estimated

In real datasets, the manifold is unknown, so geodesic distance is usually approximated from a nearest-neighbor graph:

  • Connect each point to its k nearest neighbors or points within a radius.
  • Use edge lengths based on local Euclidean distances.
  • Compute shortest paths through that graph with algorithms like Dijkstra’s algorithm or Floyd–Warshall.

This is the core idea behind Isomap, available in libraries such as scikit-learn as sklearn.manifold.Isomap.

Why it matters

Geodesic distance helps preserve the true large-scale shape of curved data when reducing dimensions. If you ignore it, points that are far apart along the manifold can look falsely close, which distorts embeddings, clusters, and visualizations. For example:

  • In image pose data, two pictures can look nearby in pixel space but represent very different positions along a motion path.
  • In customer behavior data, a nonlinear progression of engagement can be better captured by neighborhood paths than by straight-line distance.

That makes geodesic distance essential when the structure you care about is curved, connected, and local neighborhoods are more trustworthy than global straight-line measurements.

Geodesic Distance is the shortest path between two points measured along the surface or manifold that the data lies on, rather than straight through the surrounding space. In manifold learning, it captures the data’s intrinsic geometry when Euclidean distance is misleading on curved structures. This matters because methods such as Isomap rely on geodesic distances to preserve global relationships and recover meaningful low-dimensional embeddings.

Imagine traveling across the Earth. If you want to get from one city to another, the meaningful distance is not a straight line through the planet, but the path along the surface. That surface-following distance is the idea behind Geodesic Distance.

In AI, data can lie on a curved shape, even if it sits inside a huge, messy space. Geodesic Distance means measuring how far two data points are by following the shape the data naturally forms, rather than cutting straight across empty space. This matters because the “along-the-surface” route often matches the true relationship better. It helps AI keep the real structure of complex data, like poses of a moving body or different views of the same face.