I'm currently working on a variant of a non-convex low-rank matrix completion algorithm, whereby we take a uniform sample of entries in a (symmetric) matrix and look to complete said matrix. For various reasons, we're interested in trying to reduce the bandwidth of our matrix initialization for our algorithm, and our first idea was using the reverse Cuthill-McKee algorithm to do this (mostly just because it's pre-built in matlab). However, the output structure of the Cuthill-McKee algorithm provides a poor initialization for our algorithm.
I'm mostly just curious if there are other bandwidth reducing algorithms that people use regularly. I did a brief literature search and found some papers using neural networks to try and learn permutations to decrease the bandwidth, which looks interesting but is more work than I think is worth for this particular problem.