There is relatively little difference between IR and IF at an abstract level, since both are concerned with the task of getting information to people who need it. However there are certain aspects of the general problem that have been ignored by the IR literature but are especially relevant in filtering contexts. Learning and adaptation is an issue of significance in the filtering context that has not been emphasized as much in IR research.
IR has mostly dealt with the learning issue through relevance feedback. Relevance feedback refers to the reformulation of a search query in response to feedback provided by the user for the results of previous versions of the query. Work on Relevance Feedback methods in IR has a long history [38]. It has been used for the vector space model [37] and found to significantly improve performance [40]. Another approach to modeling adaptation is the use of genetic algorithms (GA) to devise adaptive algorithms for information retrieval. Yang and Korfhage [47] evolve a population of query individuals to optimize the search. Gordon [14], on the other hand, uses a method where competing representations are associated with documents. The representations are then altered over time using GA.
Learning and adaptation is, however, of much greater importance in filtering contexts. Filtering is concerned with repeated use of the system by users with long term interests. Filtering systems are often used by larger communities of people, a large number of whom might not be highly motivated information seekers [7]. Interests may not always be well defined or might not always be well expressed. In addition, the users' interests cannot be assumed to be constant, as mentioned before. Filtering systems must therefore be responsive to dynamic user interests.
Baclace [2] proposes a hybrid algorithm to evolve agents for information intake filtering. Information intake filtering (IIF) refers to ``prioritizing objects in a conceptual in-basket''. The IIF agents learn using a combination of a genetic and economic algorithm.
A number of different techniques have been used for modeling the user. Doppelgänger [35] is a user modeling system that acquires information through many sensors. For example, the ``badge sensor'' transmits physical location of the user while a ``login sensor'' tracks when people log in to computers. It uses prediction techniques such as time series analysis and Hidden Markov models to make inferences about user behavior. Any client application can obtain these inferences by making a request to the user model.