bayesian nets and hci

via green hat journal:

Eric Horvitz gave a talk on recent work at Microsoft Research's Adaptive Systems Group. First, he demonstrated a system that determines whether it should interrupt you with new information (email, IM, stock quote, etc.) by watching the applications you use, listening to the noise in the room, watching the pose of your head, and 40 other variables...

All these systems use Bayesian networks to aggregate all the variables into a probability. For the first system, it determines the probability that you would want some info by weighing the importance of the info against how busy you appear to be. I was surprised that the most important variables was so simple: if you are switching between apps frequently, you're probably not too busy. Some people suggested adding some context aware clues, like "if I'm using Visual Studio then I'm programming and don't want to be disturbed"...