Most present filtering systems are examples of what Malone has termed a cognitive filter.[18] These filters attempt to select articles for a user to view by determining the value of the information contained in the article to the user. These systems are primarily designed for use by a single user. All pieces of information that might be displayed to the user are first sent through a filter that applies a set of criteria to each piece of information. Articles which meet a sufficient number of the filter's criteria are passed on to the user with the remaining articles being discarded.
Cognitive filters can be roughly broken down into the two classes of user-profile based or rule based filters. The difference between the two lies in how the criteria used for filtering are created. In a user-profile based system, the filtering criteria are usually generated automatically in response to user actions. Many different methods have been developed for storing the filter criteria used in a user-profile system. In contrast, users explicitly enter the rules which are used to filter information in a rule based filter system. Such rules often take the form of boolean or structured query language (SQL) expressions. We will now examine some examples of each type of filtering system.
User-profile based systems attempt to extract patterns from the observed behavior of the user, and then predict which other items from the information stream will be selected or rejected based on the patterns. An example of such systems is the LyricTime project by Shoshana Loeb which compiles a profile of musical tastes for each user of an on-line jukebox.[16] LyricTime uses the user's profile to filter songs from a collection of available music. As the user accepts and rejects songs suggested by the system, the system refines the user's profile with the goal of delivering only songs which the user will approve of. LyricTime stores a user's profile as a series of keywords to look for and a confidence rating to describe how sure the system is that the user approves of songs with the keyword.
Another user-profile based filtering system is the Evolving Agent by Beerud Sheth and Pattie Maes.[26] This system attempts to learn how to filter Net News for a user by creating an new agent to filter for each topic that interests the user. These agents store information about the user's interest profile as a genotype consisting of keywords to look for and associated numerical weights. Articles are filtered by assigning each article a score proportional to the number of keywords it contains scaled by the weights of those keywords. Only articles with a suitably high score are shown to the user. The system improves its ability to filter Net News by using a genetic algorithm to breed the agents and create a new crop of agents better adapted to filtering for the user. Relevance feedback from the user is used as the fitness function which determines which agents survive and reproduce.
INFOSCOPE is an interesting hybrid news reader by Gerhard Fischer and Curt Stevens.[9] Rather than seeing itself as a sieve style filter, INFOSCOPE frames the filtering problem as one of restructuring the information space on a user by user basis to place all the articles relevant to a user in a few accessible locations. INFOSCOPE performs this restructuring by binning new articles into baskets set up by the user to contain the articles relevant to the user's interests. These baskets, called ``virtual newsgroups,'' can draw articles from many Usenet newsgroups, selecting which articles to include in the virtual newsgroup by using either rule or profile based sieve filtering techniques.