We spider using random graph walks, and it makes sure that some
x users in
y directions and
z steps away from you are spidered. However, this was the first time we'd tried to run the engine on an "incomplete" (meaning that we don't have the whole graph) data set and there are some quirks in the recommendations. It was hard to tweak the algorithm, because until there was a bunch of data, we weren't sure what tweaks would be necessary.
The skewing happens because we first check those x and y to see if we've already got enough data before running the spider. If we happen to have already spidered heavily in one subset of your friends, then you'll see the results skewed towards those until we end up spidering the others.
The other skewing, which other works better on complete graphs is that it tries to do some weighting based on how many incoming and outgoing edges users have (friends, followers), which if your tech friends have a whole lot of friends and followers and you don't, might be part of why they'd be less highly rated.
For the moment we're not really trying to compete with Mr. Tweet (though, when I tried their recommendations they were pretty bad) and whoshouldifollow (which does pretty well) -- it was just a neat way of showing off some of our engine and applying it to a different area.