Authorship Attribution of Short Messages Using Multimodal Features

In this thesis, we develop a multimodal classifier for authorship attribution
of short messages. Standard natural language processing authorship
attribution techniques are applied to a Twitter text corpus. Using character
n-gram features and a Naïve Bayes classifier, we build statistical models of
the set of authors. The social network of the selected Twitter users is
analyzed using the screen names referenced in their messages. The timestamps of the messages are used to generate a pattern-of-life model. We analyze the physical layer of a network by measuring modulation characteristics of GSM cell phones. A statistical model of each cell phone is created using a Naïve Bayes classifier. Each phone is assigned to a Twitter user, and the probability outputs of the individual classifiers are combined to show that the combination of natural-language and network-feature classifiers identifies a user to phone binding better than when the individual classifiers are used independently.