The Personalized News Filtering system is modeled as a set of Information Filtering Interface Agents. An information filtering agent assists the user with the task of finding interesting news articles in a particular domain. It has technical knowledge about the task involved, namely, information filtering. It is also aware of the interests and preferences of the user. Agents use this knowledge to automate filtering tasks for the user. Users typically have more than one filtering agent. For example, this may be the case when the user has non-overlapping news interests and would like different agents to satisfy each of these interests.
An agent is modeled as a population of profile individuals, each of which searches for articles in a small domain. While each of these profiles would typically satisfy a small part of user interests, the behavior of the population as a whole is more interesting. Together, all profiles of a population try to match the complete user interests and adapt to them. The profile contains information about where to search for articles and what kinds of articles to filter. Profiles search for articles that are similar to itself. Top-scoring articles are retrieved for presentation to the user. The articles recommended by each of the profiles are collected together and presented to the user. The user can provide positive or negative feedback for an article. User feedback has two effects. First, the profile which retrieved the article is modified based on the relevance feedback for the article. Secondly, the fitness associated with the profile increases or decreases for positive and negative feedback respectively. The fitness of a profile represents its overall suitability in serving user interests.
The population of profiles is continually evolving, in response to dynamic changes in user preferences. When the population evolves from one generation to the next, the profiles with high fitness are retained in the next generation, while the unfit ones are eliminated. The vacancies created in the population are filled in by genetic variants of the fit ones. This introduces newer members into the population, which might potentially match user interests better than the unfit profiles they replaced. If so, they will attain high fitness values and will stay in the population. Else, they will be eliminated in succeeding generations.
This chapter presents the architectural framework of a filtering agent. It describes the representation of an agent as well as the learning algorithms used. The fundamental concepts discussed include the representations for documents and profiles, the filtering algorithm and the learning mechanisms, namely relevance feedback and the genetic algorithm. Each of these is described in detail in this chapter. Implementation issues have been dealt with in the following chapter.