Manifold Learning

Nonlinear dimensionality reduction.


  • Data: input dataset


  • Transformed Data: dataset with reduced coordinates

[Manifold Learning]( is a technique which finds a non-linear manifold within the higher-dimensional space. The widget then outputs new coordinates which correspond to a two-dimensional space. Such data can be later visualized with [Scatter Plot](../visualize/ or other visualization widgets.


  1. Method for manifold learning: - [t-SNE]( - [MDS](, see also [MDS widget](../unsupervised/ - [Isomap]( - [Locally Linear Embedding]( - [Spectral Embedding](

  2. Set parameters for the method: - t-SNE (distance measures):

    • Euclidean distance

    • Manhattan

    • Chebyshev

    • Jaccard

    • Mahalanobis

    • Cosine

    • MDS (iterations and initialization): - max iterations: maximum number of optimization interactions - initialization: method for initialization of the algorithm (PCA or random)

    • Isomap: - number of neighbors

    • Locally Linear Embedding: - method:

      • number of neighbors

      • max iterations

    • Spectral Embedding: - affinity:

      • nearest neighbors

      • RFB kernel

  3. Output: the number of reduced features (components).

  4. If Apply automatically is ticked, changes will be propagated automatically. Alternatively, click Apply.

  5. Produce a report.

Manifold Learning widget produces different embeddings for high-dimensional data.


From left to right, top to bottom: t-SNE, MDS, Isomap, Locally Linear Embedding and Spectral Embedding.


Manifold Learning widget transforms high-dimensional data into a lower dimensional approximation. This makes it great for visualizing datasets with many features. We used to map 16-dimensional data onto a 2D graph. Then we used [Scatter Plot](../visualize/ to plot the embeddings.