Adaptive Information Filtering - Text Mining: Classification, Clustering, and Applications - page 206

tem, when the amount of training is extremely small or zero. 5 How should a
good filtering system learn user profiles eciently and effectively with limited
user supervision while filtering? In order to solve this problem, researchers
working on adaptive filtering have tried to develop a robust learning algorithm
that can work reasonably well when the amount of training data is small and
more effective with more training data (66) (71). Some filtering systems ex-
plore what the user likes while satisfying a user immediate information need
and trade off exploration and exploitation (75) (15). Some filtering systems
consider many aspects of a document besides relevance, such as novelty, read-
ability, and authority (70) (65). Some filtering systems use multiple forms
of evidence, such as user context and implicit feedback from the user, while
interacting with a user (70) (41).
This chapter does not cover all adaptive filtering topics in detail due to
the space limit and also because they are less “text” oriented. To finish this
section, some missed important topics are listed as follows, and the readers
are referred to the papers cited for more details
8.6.1 Beyond Bag of Words
Most of the existing adaptive filtering approaches are focused on identify-
ing relevant documents using distance measures defined in a document space
indexed by text features such as keywords. This is a very simple and limited
view of user modeling, without considering user context or other property of a
document, such as whether a document is authoritative or whether it is novel
to the user. However, even this simplest filtering task is still very hard, and
existing filtering systems do not work effectively. Bayesian graphical model-
ing, a complex data driven user modeling approach, has been used to learn
from implicit and explicit user feedback and to satisfy complex user criteria
(70).
8.6.2 Using Implicit Feedback
For most of adaptive filtering work described in this section, we assume the
system learns from explicit user feedback on whether a document delivered
is relevant or not. There is much related work on using implicit feedback in
the information retrieval community and the user modeling community. The
work in these areas can be categorized according to the behavior category
and minimum scope and have been reviewed recently (27). There are many
possible behaviors (view, listen, scroll, find, query, print, copy, paste, quote,
mark up, type, and edit) on different scope (segment, object, and class) for
system designers to choose from. Implicit feedback has also been explored
5 It is possible the system needs to begin working given a short user query and no positive
instance.