Iterative Double Clustering for Unsupervised and Semi-supervised Learning

This paper studies the Iterative Double Clustering (IDC) meta-clustering algorithm, a new extension of the recent Double Clustering (DC) method of Slonim and Tishby that exhibited impressive performance on text categorization tasks [1]. Using synthetically generated data we empirically demonstrate that whenever the DC procedure is successful in recovering some of the structure hidden in the data, the extended IDC procedure can incrementally compute a dramatically better classification, with minor additional computational resources. We demonstrate that the IDC algorithm is especially advantageous when the data exhibits high attribute noise. Our simulation results also show the effectiveness of IDC in text categorization problems. Surprisingly, this unsupervised procedure can be competitive with a (supervised) SVM trained with a small training set. Finally, we propose a natural extension of IDC for (semi-supervised) transductive learning where we are given both labeled and unlabeled examples, and present preliminary empirical results showing the plausibility of the extended method in a semi-supervised setting.