Isomap
Isometric Feature Mapping (Isomap) · Also known as: Isomap, isometric feature mapping, geodesic Isomap, nonlinear MDS
Isomap (Isometric Feature Mapping) is a manifold learning algorithm introduced by Tenenbaum, de Silva, and Langford in 2000 that discovers the intrinsic low-dimensional geometry of high-dimensional data by preserving geodesic — rather than straight-line Euclidean — distances between all pairs of points. It was one of the earliest, and most influential, nonlinear dimensionality reduction methods to demonstrate that genuinely curved data manifolds could be unfolded into a faithful low-dimensional coordinate system.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Isomap is appropriate when high-dimensional data are believed to lie on or near a smooth, low-dimensional manifold and when the goal is an interpretable global embedding that faithfully preserves the intrinsic geometric structure. It is particularly well-suited to data that form a single connected manifold without holes or highly non-convex topology. The algorithm is unsupervised and makes no parametric assumptions about the data distribution, but it does assume that the manifold is isometric to a convex subset of Euclidean space — meaning the underlying surface should be well-approximated by a flat sheet, not a torus or sphere. Sample sizes of at least 50 are typically needed to build a reliable neighborhood graph, and very high-dimensional data should first be lightly pre-reduced (e.g., with PCA) before applying Isomap. Isomap is less appropriate when data contain multiple disconnected clusters, when the manifold has holes or loops, or when computational cost on large datasets (O(n^2) memory for the distance matrix) is prohibitive.
Strengths & limitations
- Provides a globally consistent low-dimensional embedding by respecting geodesic rather than Euclidean distances, correctly unfolding curved manifolds.
- Grounded in classical MDS theory, giving a principled and mathematically transparent derivation.
- The number of retained dimensions d can be chosen objectively using a residual-variance plot.
- No parametric assumptions about the data distribution are required.
- One of the most widely cited and historically significant manifold learning algorithms, with extensive literature support.
- Requires O(n^2) memory to store the full geodesic distance matrix, making it impractical for very large datasets without landmark-based approximations.
- Sensitive to the neighborhood parameter k (or epsilon): too small yields a disconnected graph; too large short-circuits the manifold by adding spurious long edges.
- Fails on manifolds with holes, loops, or non-convex topology because shortest paths may not approximate true geodesics in those cases.
- The embedding is not parametric — new out-of-sample points cannot be projected without rerunning the full algorithm (though Nyström extension offers a partial remedy).
- Performance degrades in the presence of high noise or outliers that disrupt the neighborhood graph.
Frequently asked
How do I choose the neighborhood parameter k?
A common heuristic is to try several values (e.g., 5, 10, 15, 20) and select the smallest k for which the neighborhood graph remains fully connected. Plotting residual variance against k can also reveal a stable region. Values that are too small produce a disconnected graph; values too large introduce short-circuit edges that distort geodesic estimates.
How do I choose the number of output dimensions d?
Compute the residual variance — the correlation between geodesic distances and Euclidean distances in the embedding — as a function of d. The value at which the residual variance curve levels off (an 'elbow') indicates the intrinsic dimensionality. For visualization, d = 2 or 3 is typically chosen regardless.
Can Isomap handle out-of-sample points?
The basic algorithm requires rerunning the full geodesic computation when new points arrive. The Nyström extension provides an approximation that projects new points into the existing embedding without refitting, at the cost of some accuracy.
How does Isomap differ from t-SNE and UMAP?
Isomap is designed to produce a globally faithful embedding that preserves geodesic distances across the entire manifold, making it suitable for tasks where global structure matters. t-SNE and UMAP prioritize local neighborhood preservation and produce visualizations that cluster well but distort global distances; they are generally preferred for exploratory cluster visualization. Isomap is also fully deterministic (given k and d), whereas t-SNE has a stochastic component.
Sources
- Tenenbaum, J. B., de Silva, V. & Langford, J. C. (2000). A global geometric framework for nonlinear dimensionality reduction. Science, 290(5500), 2319–2323. DOI: 10.1126/science.290.5500.2319 ↗
- Hastie, T., Tibshirani, R. & Friedman, J. (2009). The Elements of Statistical Learning (2nd ed., Ch. 14). Springer. ISBN: 978-0-387-84857-0
- van der Maaten, L., Postma, E. & van den Herik, J. (2009). Dimensionality reduction: A comparative review. Journal of Machine Learning Research, 10, 66–71. link ↗
How to cite this page
ScholarGate. (2026, June 3). Isometric Feature Mapping (Isomap). ScholarGate. https://scholargate.app/en/machine-learning/isomap
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Kernel PCAMachine learning↔ compare
- Principal Component AnalysisMachine learning↔ compare
- t-SNEMachine learning↔ compare