I have a large text dataset clusterized. Each cluster is represented by a centroid of the vectorized texts that belong to it, the number of texts, the created date, and other parameters. I can't plot the clusters in an n-dimensional space. Which options do I have?
- Locally Linear Embedding (LLE): This method construct a set of local geometric patches on each of which a data point is reconstructed through the weighted sum of its K nearest neighbor and maps these patches into a lower dimensional space. Find the code here and I strongly advice to use both LLE and Modified LLE and use the better one (visually).
- t-SNE: Maps the similarity of points in high-dimension, into a low dimensional manifold extracted by their distance in t-distribution. Be careful to tune parameters properly.
- Spectral Embedding: Spectral Clustering in fact (or to be precise, Spectral Clustering is indeed a simple clustering on Spectrally Embedded version of data). It projects the data on eigenvectors of its Laplacian according to the magnitude of corresponding eigenvalues.
an many more ...
You can use a dimensionality reduction algorithm (like principal component analysis) to reduce the number of dimensions of the data to 2 or 3, and then perform scatter plots using the reduced variables, coloring them according to the cluster they belong to. In this blogpost a similar thing is done.