
 Description:
Stochastic neighbor embedding aims at the reconstruction of given distance, dissimilarity, or score neighborhood relations in a lowdimensional Euclidean space. This can be regarded as general approach to multidimensional scaling, but the reconstruction is based on the definition of input (and output) neighborhood probability alone. The present implementation makes use of quasi 2nd order gradientbased (l)BFGS optimization.
Neighbor relationships in the embedding space ('scatter plots') are estimated as probabilities of Gaussian or Studentt distributions; probabilities can be derived from from Gaussians over pairwise Euclidean input distances, or, as in the present case, by novel score neighborhood probabilities that do not require extra settings such as 'Gaussian width' or 'perplexity'.
The functionality is {SNE, tSNE} x {symmetric, asymmetric} where SNE is ordinary (yet wellworking) stochastic neighbor embedding, tSNE tries to avoid the 'crowding' problem by using Studentt rather than Gaussian neighborhood density assumption on the output space. Symmetric refers to forcing symmetric neighborhood relationships (as originally proposed for tSNE) for visually appealing plots, else asymmetric relationship reconstruction might be preferred for better representation qualities of the embedded point cloud. In the original publication of tSNE, symmetric also includes neighborhood probability inference for all points rather than the sum of separate pointwise probabilities. That original assumption is not used in this package, because pointwise reconstruction is supposed to be more specific than for general neighborhood probability distributions in the input and output space.
As additional feature, the embedding quality of data points is assessed by the contributions of embedding point placement to the cost function, i.e. the sum of absolute KLdivergence gradients caused by individual points.
Acknowledgements: I. The work on SNE and tSNE I is highly appreciated "Visualizing Data using tSNE", JMLR 9, pp. 25792605, 2008, and the freely available implementations by Laurens van der Maaten. II. The great (l)BFGS optimizer (fminlbfgs.m) of DirkJan Kroon found at http://www.mathworks.de/matlabcentral/fileexchange/23245 included here is STRONGLY acknowledged.
 Changes to previous version:
Negligible changes for consolidating the code.
 BibTeX Entry: Download
 Supported Operating Systems: Platform Independent
 Data Formats: Matlab
 Tags: Dimension Reduction, Mds, Multidimensional Scaling, Sne, Stochastic Neighbor Embedding
 Archive: download here
Comments
No one has posted any comments yet. Perhaps you'd like to be the first?
Leave a comment
You must be logged in to post comments.