Waern (1997): Local plan recognition

Many user models try to interpret user actions and fit them to know tasks

Two types of plan recognition:

'Keyhole' plan recognition (background, silent monitoring)
Intended plan recognition (collaborative dialogue)

Addressing keyhole recognition:

A problem with all active user modeling - the information obtained is of low quality

Users often see applications as dumb tools

Illustrated with a news filtering application

Manual marking of what is 'interesting' or not can be effective,
But, of course, is intrusive

Automatic recognition:

Inference of user behavior
But how to recognize what is interesting and what isn't?