Next: Previous Work Up: No Title Previous: Overview of this

Personalized Information Filtering

A tremendous amount of news and information is created and delivered over electronic media. This has made it increasingly difficult for individuals to control and effectively manage the potentially infinite flow of information. Ironically, just as more and more users are getting online, it is getting increasingly difficult to find information unless one knows exactly where to get it from and how to get it. Tools to regulate the flow are urgently needed to prevent computer users from being drowned by the flood of incoming information.

Information filtering systems can help users by eliminating the irrelevant information and by bringing the relevant information to the user's attention. Filters are mediators between the sources of information and their end-users. Belkin and Croft [7] provide a good description of Information filtering (IF) and identify the similarities and differences with Information Retrieval (IR). Filtering contexts typically involve a dynamic stream of information, as opposed to static data bases used in traditional IR systems. Due to the dynamic nature of the stream, timeliness of information assumes added significance in the filtering context. The amount of data involved in filtering environments is usually very large and unstructured.

IF involves repeated interactions over multiple sessions with users having long-term goals. This is in contrast to IR systems, where the user typically has a short term information need that is satisfied within a single session. This implies that an IF system must remember the user and individualize its performance for her. Filtering systems maintain profiles which are representations of the user interests. Privacy of the user profiles is an issue that is of concern in filtering systems. Adequate precautions must be taken to ensure that complete control over the user profile rests with the user.

Filtering systems are much more likely to be used by a wider community of users than IR systems. The users of IF might not be highly motivated in their information seeking and may not have well defined interests. Therefore, IF systems need to be much more user-friendly than IR systems which have been mostly aimed at the highly motivated information seeker with very specific information needs.

A Personalized IF system should be highly responsive to the needs of the user. Since a filtering system involves repeated interactions with the user, the system should get a better feel for the user's needs over time. Assuming that a lot of the user actions are consistent, the system should get increasingly better at matching her needs over time. The system should be able to gradually converge to that part of user needs which is predictable and consistent. Furthermore, since the interaction could extend over a long period of time, the user's interests cannot be assumed to stay constant. The change in interest could be anything from a slight shift in relative priorities to completely losing interest in some domain and gaining interest in another. The system must be able to detect or must allow the user to indicate the change in interests and should respond by adapting to these changes. Finally, considering the possibility that the information seeker might not be highly motivated, the system should be capable of recommending new potentially interesting information based on what it already knows about her. The system must be able to explore newer domains and prospect for interesting information.

To summarize, the personalized information filtering problem is to design a system that can provide information that matches user needs in a consistent and timely manner. The system must also accommodate changes in user needs and adapt to those changes. Ideally, the system should be capable of entertaining not only the currently known needs of the user, but also exploring different domains to find articles of potential user interest. Thus, designing a system that is specialized, adaptive and exploratory is the ``holy grail'' of personalized information filtering.




Next: Previous Work Up: No Title Previous: Overview of this


MIT Media Lab - Autonomous Agents Group - agentmaster@media.mit.edu