Evaluation of text clustering methods and their dataspace embeddings: an exploration

Fair evaluation of text clustering methods needs to clarify the relations between 1)pre-processing, resulting in raw term occurrence vectors, 2)data transformation, and 3)method in the strict sense. We have tried to empirically compare a dozen well-known methods and variants in a protocol crossing three contrasted open-access corpora in a few tens transformed dataspaces. We compared the resulting clusterings to their supposed "ground-truth" classes by means of four usual indices. The results show both a confirmation of well-established implicit combinations, and good performances of unexpected combinations, mostly in spectral or kernel dataspaces. The rich material resulting from these some 450 runs includes a wealth of intriguing facts, which needs further research on the specificities of text corpora in relation to methods and dataspaces.