Research Article
James Tesiero
Abstract
In this work, we discover groups of similar users of a topic recommendation system. The data sets used are from Tipster Newsgroups, which are used in the annual TREC (Text Retrieval Conference) competition. The users are simulated, represented by tag/rating pairs. The documents in the Tipster data sets are clustered with a topic clustering algorithm that is a subset of the topic recommendation system being proposed in this paper. The users query the clusters derived from the topic clustering algorithm with tags, then rate the degree of relevance the content returned by the system has to their tag. It is shown in this work that, starting from a random sampling of the clusters by the users, and a random initial distribution of ratings per user, that a topic recommendation engine powered by a Boltzmann machine with nearest neighbor interactions results in two distinct clusters of users: those that converge quickly to a particular single topic, and those who explore a few different topics in a way that is periodic in time. This allows new users entering the system to be clustered and hence given a more relevant experience earlier in the process.