I wonder if it makes sense for GitHub to use graph-theoretic measures like PageRank instead of raw stars. In simple terms, a repo is considered important if it is starred or forked by GitHub users who maintain other important repos.
It’s more expensive to compute, but the resulting scores would be more trustworthy unless I’m missing something.
That sounds closer to achieving a good outcome. Of course I think anything that includes the set of all users as columns will be game-able. You need to either choose the set yourself from "trusted peers" or "foaf" degrees, or maybe better use retroactive signals rather than purely like-driven approaches.