Unsupervised Classification of Text-Centric XML Document Collections

This paper addresses the problem of the unsupervised classification of text-centric XML documents. In the context of the INEX mining track 2006, we present methods to exploit the inherent structural information of XML documents in the document clustering process. Using the k-means algorithm, we have experimented with a couple of feature sets, to discover that a promising direction is to use structural information as a preliminary means to detect and put aside structural outliers. The improvement of the semantic-wise quality of clustering is significantly higher through this approach than through a combination of the structural and textual feature sets. The paper also discusses the problem of the evaluation of XML clustering. Currently, in the INEX mining track, XML clustering techniques are evaluated against semantic categories. We believe there is a mismatch between the task (to exploit the document structure) and the evaluation, which disregards structural aspects. An illustration of this fact is that, over all the clustering track submissions, our text-based runs obtained the 1st rank (Wikipedia collection, out of 7) and 2nd rank (IEEE collection, out of 13).