In order to explore the potential of machine learning methods to automatically
acquire search control knowledge for WebWatcher, we collected a set of data
from 30 sessions using WebWatcher to search for technical papers. In each
session the user began at the web page shown in Figure 1, and
searched for a particular type of technical paper following links forward from
there. Searches were conducted by three different users. The average depth
of a search was 6 steps, with 23 of the 30 searches successfully locating a
paper. Each search session provided a set of training examples corresponding
to all the 
pairs occurring on each page visited by the
user.