Efficient Unsupervised Authorship Clustering Using Impostor Similarity

Some real-world authorship analysis applications require techniques that scale to thousands of documents with little or no a priori information about the number of candidate authors. While there is extensive research on identifying authors given a small set of candidates and ample training data, almost none is based on real-world applications of clustering documents by authorship, independent of prior knowledge of the authors. We adapt the impostor method of authorship verification to authorship clustering using agglomerative clustering and, for efficiency, locality sensitive hashing. We validate our methods on a publicly available blog corpus that has been used in previous authorship research. We show that our efficient method matches previous results for authorship verification on the blog corpus. We extend previous results to show that the impostor algorithm is robust to incorrectly selected impostors. On the authorship clustering task, we show that the impostor similarity method clearly outperforms other techniques on the blog corpus.