A model news filtering economy

Our model system is an information filtering economy, first described in ref. [6]. It consists of a source agent that publishes news articles, C consumer agents that want to buy articles they are interested in, B broker agents that buy selected articles from the source and resell them to consumers, and a system infrastructure that provides communication and computation services to all agents. A diagram showing part of our model system is shown in Fig. 1. The ellipse at the top represents the source agent, brokers are in the middle, and consumers are at the bottom. Each agent's internal parameters (defined below) are printed inside its ellipse. The system is represented by the rectangle on the left. Solid lines represent the propagation of a sample article through broker 1. Broken lines indicate payment, and are labeled with symbols (explained below) for the amount paid.

  
Figure 1: Part of an idealized news filtering economy. Only a subset of agents is shown. See text for interpretation of symbols.

The source agent publishes one article at each time step t, and waits until that article has propagated through the system before publishing the next. It classifies articles according to its own internal categorization scheme, assigning each a category index j when it is offered. The nature of the categories, and the number J of them, do not change. We represent this (hidden) classification scheme by a random process in which an article is assigned category j with fixed probability . The set of all is the source's category prevalence vector . Each article labeled with its category index and offered for sale to all brokers at a fixed price . For each article sold to each broker, the source pays a fixed transport cost .

Upon receiving an offer, each broker b decides whether or not to buy the article using its own evaluation method to select which categories it is "interested" in. The broker's evaluation method--which may involve a categorization scheme entirely different from that used by the source--is approximated by an interest vector , where represents the probability for b to purchase an article labeled with category j. Analysis of the model [6] shows that it is in broker b's best interest to set the individually to either 0 or 1.

When broker b purchases an article, it immediately sends it to a set of subscribing consumers, paying transportation cost for each. Subscribers may examine the article, but must pay the broker if they want the right to use ("consume") it. The broker's internal parameters and are under its direct control.

Subscriptions are represented by a subscription matrix S, where if consumer c subscribes to broker b, and if not. Subscriptions are maintained only with the consent of both parties and may be cancelled by either. For example, a broker b might not wish c to subscribe if the cost of sending articles exceeds the expected payment from c, or c might not find it worthwhile to subscribe to b if the cost of sifting through lots of junk outweighs the benefit of receiving the rare interesting article. The requirement that the agreement be bilateral is represented by setting , where if broker b wants consumer c as a subscriber and if not; similarly, if consumer c wants to subscribe to broker b and if not.

Each consumer waits for articles to arrive from the brokers it subscribes to. When a consumer receives one or more copies of an article, it pays the computation cost to evaluate whether it is interested in the article, then decides whether (and from whom) to buy it. Like the brokers, the consumers' evaluation function is approximated by a stochastic process parametrized by an interest vector : consumer c will be interested in an article labeled with category j with fixed probability . If a consumer is interested in an article, it then selects from the set of brokers it subscribes to the one broker with the most attractive offer; we shall assume the most attractive offer is the cheapest one. The consumer then decides whether its interest justifies paying for that article. For reasons of simplicity, we model this decision process as follows: each consumer assigns a global constant anticipated value V to each article it is interested in. Then if , it purchases the usage rights; otherwise it discards the article unused.

Each broker's or consumer's decision-making process may be expressed as an attempt to optimize its utility function, defined as the amount of net "value" or "utility"--however it may be measured--gained by making that particular decision. In the system described here, the expected utility per article for each broker and consumer may be explicitly formulated from the system variables. For consumers, the anticipated value V provides the fundamental benchmark for measuring utility. For brokers, the appropriate measure of utility is profit, defined in the usual way as revenue less expenses. General expressions for consumer utility and broker profit may be found in ref. [6].

In closing this section, let us observe that our model disregards a number of potentially interesting features of real information filtering systems. By assigning a single category index to each article, it disregards articles that fall into multiple categories. By having brokers and consumers evaluate each article independently of previous articles, it disregards the possibility of multiple articles containing redundant information. By assigning a constant value to each article purchased by each consumer, it disregards differences among consumers and among articles. By approximating the brokers' and the consumers' evaluation methods by stochastic functions, it disregards the advantages and difficulties associated with automatic evaluation of articles' semantic content.

Each of these represents a refinement of the model which brings it closer to a real information filtering system. Nevertheless, they lie outside the scope of the present article. As we noted in the introduction, the focus of this article is the economic properties of the system, as opposed to the purely informational ones.