This depends on how the similarity is measured. If it is measured as % of agreement between comment texts then it's more likely to have better agreement with so

This depends on how the similarity is measured. If it is measured as % of agreement between comment texts then it's more likely to have better agreement with someone who has fewer comments rather than more. For example I suspect that if we generated 1000 random users with random gibberish comments and varied their comment numbers from 1 to 10 or so, the top similarities would be biased towards low comment-count random users. This would be because having one randomly generated comment match your style is easier than having 2 randomly generated comments match the same style. And if that's the case then the same issue would transfer to comparing real users. @busymom0 suggested a great solution in this thread - only do comparisons based on "n" (like last 50) comments. This way every similarity would be measured using the same number of comments and users with low comment counts would be excluded automatically.