Applying SVD + PPMI to Image Tags

[BradKML]

As observed in https://www.aclweb.org/anthology/L18-1156.pdf the same methods can be applied to image metadata.
https://github.com/ThomasAger/sparse_ppmi/blob/master/sparse_ppmi.py

If one were to provide a list of image metadata JSON files, with tag IDs, then how could one go about constructing a sparse-like Co-Occurance matrix? And from that, a sparse-like PPMI Matrix for SVD training?

[BradKML]

Also oofing on https://github.com/jfilter/hyperhyper