Similarity-Based Text Clustering: A Comparative Study

Summary

Clustering of text documents enables unsupervised categorization and facilitates browsing and search. Any clustering method has to embed the objects to be clustered in a suitable representational space that provides a measure of (dis)similarity between any pair of objects. While several clustering methods and the associated similarity measures have been proposed in the past for text clustering, there is no systematic comparative study of the impact of similarity measures on the quality of document clusters, possibly because most popular cost criteria for evaluating cluster quality do not readily translate across qualitatively different measures. This chapter compares popular similarity measures (Euclidean, cosine, Pearson correlation, extended Jaccard) in conjunction with several clustering techniques (random, self-organizing feature map, hypergraph partitioning, generalized k-means, weighted graph partitioning), on a variety of high dimension sparse vector data sets representing text documents as bags of words. Performance is measured based on mutual information with a human-imposed classification. Our key findings are that in the quasiorthogonal space of word frequencies: (i) Cosine, correlation, and extended Jaccard similarities perform comparably; (ii) Euclidean distances do not work well; (iii) Graph partitioning tends to be superior especially when balanced clusters are desired; (iv) Performance curves generally do not cross.