Locally Linear Embedding (LLE)
Also known as: LLE, manifold learning, nonlinear dimensionality reduction, yerel doğrusal gömme
Locally linear embedding, introduced by Sam Roweis and Lawrence Saul in 2000, is a manifold-learning method for nonlinear dimensionality reduction. It assumes that although data may curve through a high-dimensional space, each point and its neighbours lie approximately on a flat patch. LLE captures each point as a weighted combination of its neighbours and then finds a low-dimensional layout that preserves those same local relationships, unrolling curved structure into a faithful low-dimensional map.
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
Use LLE to visualize or reduce the dimensionality of data believed to lie on a smooth nonlinear manifold — for example, images under continuous variation (pose, lighting), motion-capture trajectories, or other data where a few latent factors generate high-dimensional observations. It is a classic alternative to Isomap (which preserves geodesic distances) and a precursor to t-SNE and UMAP. LLE is sensitive to the neighbourhood size k and to noise and non-uniform sampling, can struggle when the manifold is not well sampled or has holes, and does not provide a natural way to embed new points out-of-sample. For general-purpose visualization of clusters, t-SNE or UMAP are now usually preferred; LLE remains valuable conceptually and when local linearity genuinely holds.
Strengths & limitations
- Captures nonlinear manifold structure that linear methods like PCA miss.
- Reduces to a single sparse eigenproblem with no local minima or iterative tuning of an objective.
- Neighbourhood-reconstruction weights are invariant to local rotation and scaling.
- Conceptually elegant, with relatively few parameters (mainly the neighbourhood size).
- Sensitive to the number of neighbours k and to noise and uneven sampling.
- Struggles with manifolds that are poorly sampled, have holes, or vary in density.
- No straightforward out-of-sample extension for embedding new points.
- Often outperformed by t-SNE/UMAP for cluster visualization in practice.
Frequently asked
How is LLE different from PCA?
PCA finds a single global linear subspace, so it cannot unfold curved (nonlinear) structure. LLE assumes the data lie on a nonlinear manifold that is locally flat and preserves each point's local neighbourhood relationships, allowing it to flatten curved manifolds that PCA would distort.
How does LLE compare to Isomap?
Both are early manifold-learning methods. Isomap preserves global geodesic distances along the manifold via shortest paths on a neighbourhood graph, while LLE preserves only local reconstruction weights. LLE is computationally lighter and more local in focus; Isomap better preserves global geometry when the manifold is convex and well sampled.
Why does the choice of k matter so much?
k sets the size of the neighbourhood assumed to be linear. Too few neighbours fragment the manifold and make the embedding unstable; too many include points where the manifold has curved away, violating local linearity. Results can change noticeably with k, so it should be tuned and assessed.
Sources
- Roweis, S. T., & Saul, L. K. (2000). Nonlinear dimensionality reduction by locally linear embedding. Science, 290(5500), 2323–2326. DOI: 10.1126/science.290.5500.2323 ↗
How to cite this page
ScholarGate. (2026, June 2). Locally Linear Embedding (LLE). ScholarGate. https://scholargate.app/en/machine-learning/locally-linear-embedding
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.
Compare side by side →