Isomap
Non-linear dimensionality reduction preserving geodesic (manifold) distances between points
Isomap
Non-linear dimensionality reduction preserving geodesic (manifold) distances between points.
When to use:
- Data lies on a non-linear manifold
- Want to preserve global distances
- Need interpretable low-dimensional representation
- Moderate-sized datasets
Strengths: Preserves global geometry, interpretable, can transform new data Weaknesses: Sensitive to neighborhood size, fails if manifold is not well-sampled, slow on large data
Model Parameters
N Components (default: 2, required) Embedding dimensions.
N Neighbors (default: 5) Number of neighbors for graph construction.
- Small (3-5): Captures fine local structure
- Medium (5-15): Balanced (default)
- Large (20+): More global, but may shortcut the manifold
Metric (default: "minkowski") Distance metric:
- minkowski: Generalized (default)
- euclidean: Standard distance
- manhattan: L1 distance