Unsupervised document clustering using multi-resolution latent semantic density analysis

To find meaningful groupings in a given document collection, it is essential to learn the right granularity for the domain, uncover core themes and attendant outliers, and derive suitable labels with which to characterize each of the resulting clusters. The outcome is therefore affected both by the choice of representation and by the behavior of the clustering algorithm. This paper advocates a strategy which combines density-based clustering with latent semantic feature extraction. Documents are first mapped into a latent semantic vector space, and then clustered in that space on the basis of a multi-resolution density measure. Empirical evidence gathered on several document collections suggests that this procedure is effective in identifying semantically sound document clusters.