Comparing compression models for authorship attribution

In this paper we compare different compression models for authorship attribution. To this end, three different types of compressors, Lempel-Ziv type (GZip), block sorting type (BZip) and statistical type (PPM), along with two different similarity measures were considered in our experiments. Besides, two different attribution methods are analyzed in this paper. Through a series of experiments performed on two different databases, we were able to show that all the compressors behave similarly, but the similarity measures can vary considerably depending on the strategy used for authorship attribution. Our results corroborate with the literature in the sense that compression models are a good alternative for authorship attribution surpassing traditional pattern recognition systems based on classifiers and feature extraction.