I think that would have a stronger tendency to create an echo chamber.

If anything, I'd suggest the reverse: it should count for more to have more votes from a less tightly coupled portion of the graph. As in, cluster the set of users based on various types of interactions, and include a multiplicative factor in scoring based on the distinctness of the users upvoting something. That way, you can get a modest score by appealing strongly to a small subset of users, but a massive score only by appealing to many distinct sets of users.