Using the αβ-Neighborhood for Adaptive Document Filtering

In this paper, we address the problem of adaptive document filtering. Traditionally, user profiles are represented by the centroid of the available examples, assuming that these are homogeneously distributed around this centroid. However, these examples may be irregularly distributed, being some areas more populated than others. While, in this case, the homogeneity assumption may not be globally true, it may still hold locally. In order to handle this phenomenon, we introduce a new approach in which a binary classifier for each user profile is used and more than one document is considered in the classification task. To decide whether a new document is relevant to the user or not, our approach uses a Nearest Neighbor classifier based on a neighborhood which inspects a sufficiently small area surrounding the new document. Experiments carried out on the TREC-11 collection show the effectiveness of the proposed method.