cmdscale in R is pretty slow. It takes about 1.5 hours to do multidimensional scaling on 10000 points.
If you operate with social network data it could be useful to try http://www.vosviewer.com/... Well, its not universal tool, but it is adapted to network data :)
Olesya, I'm trying to use Vosviewer to visualize my data. My plan is to not specify the x and y coordinates of the map but let the program construct them from the network file. However, the program keeps ask me to specify x and y coordinates in the map file. Do you know how to fix this? Thanks.
I have worked with both Matlab and from the R package. You do not need to specify x-y in either . You can try that if you wish.
One approximate way that has worked well for me is to 1) reduce the sample size with a vector quantization algorithm (use, e.g. a self organizing map or Neural gas to get, say, 500 prototypes), 2) project the prototypes to obtain 500 projected points and finally 3) learn the mapping between the 500 input data space and their 500 2D projections using e.g. ANN (e.g. RBFN) or other efficient interpolation methods and use it to map your 10000 data points.
If it is intended for visualization the method is far much faster, can deal with even more points, and can yield a rather good nonlinear DR projection.
.
same suggestion as Ignacio above except I would use neural gas instead of som since in this case the only thing needed is quantization and not visualization (neural gas is less constrained than som by the topological neighbourhood relationships and gives better reconstruction errors).
moreover, the interpolation step may not be required ; it depends what is expected from mds in the course of the analysis (visualizing/analyzing the mds projection of the prototypes might be sufficient)
.
.
by the way, locally linear embedding (lle) is a nice alternative to mds ; should be rather fast for embedding in 2D
http://www.cs.nyu.edu/~roweis/lle/papers/lleintro.pdf
.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining