Introducing Biases in Document Clustering

Abstract

In this paper, we present three criteria for introducing biases in document clustering algorithms, when information characterizing the document collections is available. We focus on collections known to be the result of a document categorization or sample-based document filtering process. Our proposals rely on profiles, i.e., document samples known to have been used for obtaining the collection, to extract statistics which determine the biases to introduce. We conduct an experimental evaluation over a number of collections extracted from the widely used corpus RCV1, which allows us to confirm the validity of our proposals and determine a number of situations where biased clusterings, according to different criteria, outperform their unbiased counterparts.