What is the form of the knowledge required by WebWatcher? In general, its task is to suggest an appropriate link given the current user, goal, and web page. Hence, one general form of knowledge that would be useful corresponds to knowledge of the function:
where Page is the current web page, Goal is the information sought by the user, User is the identity of the user, and Link is one of the hyperlinks found on Page. The value of LinkUtility is the probability that following Link from Page leads along a shortest path to a page that satisfies the current Goal for the current User.
In the learning experiments reported here, we consider learning a simpler function for which training data is more readily available, and which is still of considerable practical use. This function is:
Where the value of 
is the probability that an arbitrary user
will select Link given the current Page and Goal. Notice here the
User is not an explicit input, and the function value predicts only whether
users tend to select Link -- not whether it leads optimally toward to the
goal. Notice also that information about the search trajectory by which the
user arrived at the current page is not considered.
One reason for focusing on 
in our initial experiments is that
the data automatically logged by WebWatcher provides training examples of this
function. In particular, each time the user selects a new hyperlink, a
training example is logged for each hyperlink on the current page,
corresponding to the Page, Goal, Link, and whether the user chose this
Link.