A neural network consists of a network of nodes which are composed of output and input values. When a node is activated its output value is passed along links to other nodes. An input value of a node is computed as the weighted sum of all incoming links. The weight of a link can be modified using training data.
Neural networks can be used to models the interests of a user in a web environment. The neural network is formed and modified as a result of the articles a user has read or rejected. When a word occurs often in articles a user has read a node is introduced to the network. A link between two nodes is established when their corresponding words appear often in the same article and the weight of the link is determined by the frequency in which the words appear together. The initial output value of a node is determined by the occurrences of the corresponding word in the articles. When the word occurs in an article read by the user the value is increased, when the word occurs in an article that is rejected by the user this value is decreased.
New articles are ranked by passing the words to the neural network and activating the corresponding nodes. A node connected to a set of active nodes becomes active itself when its input value is above some firing threshold. The process of node firing is repeated for several iterations until the network reaches a stable state from which no new nodes can be activated. The rank of an article is then determined by the sum of all output values of the active nodes. In practice only a limited number of iterations are allowed to restrict the computational complexity of the network.