The Content Discovery app at brainstorm.ninja demonstrates one of the simplest solutions to this problem: don’t use an Influence Score that can be greater than 1 as the weight.
At brainstorm: first, the generic Influence Scores are calculated, where all ratings are either a 1 (a follow) or a 0 (a mute). This is used to calculate the generic Influence Score, gIS.
Next, contextual Influence Scores, cIS, are calculated using contextual endorsements. An example of a context would be “Artist.” A context-specific Follow is interpreted as a score of 2, and a Superfollow as a score of 5. In this manner, a pubkey who is an artist can get a cIS in the context of “Artist” that is as high as 5. But this doesn’t result in divergence because gIS, not cIS, are used as weights when calculating weighted averages, and gIS are never higher than 1. The cIS is then used to produce the list of recommended npubs if you’re looking for Artists to follow.