NE: Neighbor Embedding for visualization [Download]

This software package implements scalable optimization of Neighbor Embedding (NE), a set of a dimensionality reduction algorithms recently developed for data visualization. An NE finds a 2D or 3D mapping of high-dimensional data, trying to preserve the neighborhood according to certain information divergences. Typical NEs include

• Stochastic Neighbor Embedding (SNE)
• Symmetric Stochastic Neighbor Embedding (s-SNE)
• Student t-Distributed Stochastic Neighbor Embedding (t-SNE)
• Elastic Embedding (EE)
• Neighbor Retrieval Visualizer (NeRV)
• Weighted t-SNE (wt-SNE)

On the right is the scatter plot of the MNIST handwritten digits. See below or in the gallery for more visualization examples by using NE.

• MNIST (70000 handwritten digit images)
• COIL-20 (1440 images of 3D-toys in 20 classes)
• CUReT (5612 texture images in 61 classes)
• WorldTrade (network data of trading amounts between 80 countries)
• USAir97 (network data of direct flights between 332 airports)
• 20 Newsgroups (20000 text documents form 20 newsgroups)
• TIMIT (510222 speech samples in 49 phoneme classes)
• PIE (1166 face images in different lighting conditions)
• Five Phonemes (4508 speech samples in 5 phoneme classes)

The usage is decribed in "readme.txt" in the package.

If you use the NE software in any publishable work, please cite at least one of the following papers:

• Zhirong Yang, Jaakko Peltonen and Samuel Kaski. Scalable Optimization of Neighbor Embedding for Visualization. In International Conference on Machine Learning (ICML2013), pages 127-135, Atlanta, USA, 2013. [pdf] [supplemental]
  
  
• Zhirong Yang, Jaakko Peltonen and Samuel Kaski. Optimization Equivalence of Divergences Improves Neighbor Embedding. In International Conference on Machine Learning (ICML2014), pages 460-468, Beijing, China 2014. [pdf] [supplemental]
  
  

This is experimental software provided as is; we welcome any comments and corrections but cannot give any guarantees about the code. If you have any comments or bug reports, please direct them to Zhirong Yang.