1
$\begingroup$

For a Random Forest, we can construct a N x N (where N is the number of data points) proximity matrix P where P[i,j] is how "close" the i-th data point is from the j-th data point. In Gilles Loupes' PhD dissertation, he shows an example of a very beautiful proximity visualization using the MNIST dataset:

My question is - how are these proximity plots made? Is there any intuitive difference between that and your traditional distance/similarity matrices? For example, if I have N = 500 data points, should I run the proximity matrix through some sort of dimensionality reduction technique like PCA / SVD / t-SNE so it is of form 500 x 2, and the visualize it?

$\endgroup$
  • $\begingroup$ Is there no description of how the plot was constructed in the text of the dissertation that you cite? $\endgroup$ – Sycorax says Reinstate Monica May 20 at 18:10
  • $\begingroup$ There is not.@Sycorax $\endgroup$ – Yu Chen May 20 at 18:10
1
$\begingroup$

A few lines later in the dissertation, the author writes

Figure 4.6 represents the proximity matrix learned for a 10-class handwritten digit classification task, as projected on a plane using Multidimensional Scaling [Kruskal, 1964].

In addition to the citation the author provides, you can find more information about by searching for posts bearing the tag.

$\endgroup$
  • $\begingroup$ Thank you, I guess I did not see this when I was perusing the paper. $\endgroup$ – Yu Chen May 20 at 19:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.