Many information search, classification, and extraction problems
are variations of a single problem in which:
each item in a potentially infinite stream of items must be classified,
that classification must be made before the next item is seen,
that classification must be made based largely
on implicit information
provided by a teacher's behavior rather than explicit guidance,
incorrect classifications are allowed as long as asymptotic behavior
is acceptable,
no classification can take arbitrarily long,
nor use arbitrary amounts of computation.
Here are several instances of this general problem:
Arrange mail or news articles
into articles a user is likely to want to see
and articles the user is not likely to want to see.
Find webpages that a user is likely to want to see.
A user is searching for a hard to describe image in a large image database;
given that the user is willing to implicitly grade
the closeness of other images to it, help the user find the image.
A user just gave the following command or page reference
to an operating system, graphical interface, or application program,
but it's not quite correct,
figure out which one was meant.
A game player used a particular rule
to generate all of the following phenomena;
guess the rule.
I've watched all these movies
(listened to all these singers, eaten all these foods, read all these books,
seen all these paintings, driven all these cars, rented all these apartments,
visited all these cities, dated all these people)
and these had property X while those didn't,
find new things likely to have property X.
Constructing a computer program general enough to be profitably used
for each of these variations with relatively small modification per variation
is the information helper problem.
