Hmm, i think its a very difficult task. I think perhaps a useful thing would be to intentionally break the bubbles. Something like an inverse-recommendation engine. E.g. instead of "people who read article X also read Y", you say "people who read X did not read Z". Or if you had access to people's facebook likes (does facebook provide an API?) you could find the most anti-correlated users and recommend each other's liked stuff etc.
I think people in general avoid the other side, and they don't have enough time for that. Thats why i suggest a mix up of sides. That way you read each story once, but as a whole you get exposed to both contexts.
