yaolubrain / DOSNES

Doubly Stochastic Neighbor Embedding on Spheres

DOSNES

DOSNES is a new method to visualize your data.

Project Page

http://yaolubrain.github.io/dosnes/

Paper

Doubly Stochastic Neighbor Embedding on Spheres
Yao Lu, Jukka Corander, Zhirong Yang
Pattern Recognition Letters, 2019

How to use?

Here is a simple example.

% Generate data and its similarity matrix
X = randn(1000,10);
D = pdist2(X,X,'squaredeuclidean');
P = exp(-D);

% Normalize the similarity matrix to be doubly stochastic by Sinkhorn-Knopp method
for i = 1:100
    P = bsxfun(@rdivide,P,sum(P,1));
    P = bsxfun(@rdivide,P,sum(P,2));
end    

% Run t-SNE with the spherical constraint
Y = tsne_p_sphere(P);

% Normalize Y to have unity radius for visualization
Y = bsxfun(@rdivide,Y,sqrt(sum(Y.^2,2)));

% Save the data 
dosnes_data = [Y ones(length(Y),1) 5*ones(length(Y),1)];
csvwrite('data.csv',dosnes_data);

Now open dosnes.html with Firefox. Don't use Chrome. You now have the DOSNES visualization in your browser.

For visualizations of more features and real world data, please see the demo folder.

In main javascript code of the demos, makeTextSprite() is to create the facing-to-viewer text labels. In the CSV files, the first three columns are XYZ-coordinates, the forth column is the class label and the last column is the size of the data points.

Python Implementation

https://github.com/Coni63/DOSNES