MathIsimple

Isomap (Isometric Mapping)

Master isometric mapping for preserving global geodesic distances. Learn how Isomap constructs neighbor graphs, computes shortest paths, and applies MDS to preserve manifold topology.

Module 7 of 9
Advanced Level
90-120 min

What is Isomap?

Isomap (Isometric Mapping) is a nonlinear dimensionality reduction method that preserves geodesic distances (shortest paths along the manifold) rather than Euclidean distances. It addresses the limitation of linear methods that compute "straight-line" distances in high-dimensional space, which don't reflect the true structure of curved manifolds.

Key Problem

On a curved manifold, the Euclidean distance (straight line) between two points may be misleading. The geodesic distance (shortest path along the manifold surface) better reflects true similarity.

Example: On Earth's surface, the geodesic distance between two cities is the great circle route, not a straight line through the Earth.

Geodesic Distance Concept

The geodesic distance between two points on a manifold is the length of the shortest path that stays on the manifold surface.

Computing Geodesic Distances

Isomap approximates geodesic distances by:

  1. 1. Constructing a neighbor graph where each sample connects to its k nearest neighbors
  2. 2. Computing shortest paths in this graph (using Dijkstra or Floyd-Warshall algorithm)
  3. 3. Using these shortest path lengths as approximations of geodesic distances

Isomap Algorithm Steps

The complete Isomap algorithm:

Step 1

Construct Neighbor Graph

For each sample xix_i, find its k nearest neighbors based on Euclidean distance. Connect xix_i to its neighbors with edges weighted by Euclidean distance. Non-neighbors have infinite distance.

Step 2

Compute Geodesic Distances

Use shortest path algorithm (Dijkstra or Floyd-Warshall) to compute shortest path distances between all pairs of samples in the neighbor graph. These approximate geodesic distances distG(xi,xj)dist_G(x_i, x_j).

Step 3

Apply MDS

Use the geodesic distance matrix as input to MDS algorithm. MDS finds low-dimensional embeddings ziz_i such thatzizjdistG(xi,xj)\|z_i - z_j\| \approx dist_G(x_i, x_j).

Advantages and Limitations

Advantages

  • • Preserves global manifold topology
  • • Handles nonlinear structures
  • • Works well for Swiss roll-type data
  • • Preserves geodesic distances

Limitations

  • • Sensitive to neighbor parameter k
  • • Sensitive to noise and outliers
  • • O(m²) complexity for shortest paths
  • • Requires connected graph