Mean User-Text Agglomeration (MUTA): Practical User Representation and Visualization for Detection of Online Influence Operations

Evan Crothers, Herna L. Viktor, Nathalie Japkowicz

Abstract: Online influence operations (OIOs) present a serious threat to the integrity of online social spaces and to real-world democratic elections. While many OIO detection approaches have focused on classification algorithms for individual social media posts (often with artificially balanced datasets), we present a novel system centering around a human analyst. This system incorporates a user representation and visualization procedure for unbalanced social media data. Our content-based social media user representation, the Mean User-Text Agglomeration (MUTA), summarizes a user’s social media activity with respect to Transformer embeddings of texts authored by the user. We apply MUTA to a real social media dataset in advance of an election event and flag a number of suspicious Reddit users that were later removed by the social media platform. When projected to a 2-dimensional visualizable space, MUTA user representations are shown, via extrinsic cluster quality measures, to outperform BERT representations for analyst identification of OIO accounts.