Learning Invariant Representations of Social Media Users

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1684–1695, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1684 Learning Invariant Representations of Social Media Users Nicholas Andrews and Marcus Bishop Human Language Technology Center of Excellence Johns Hopkins University {noa,marcus.bishop}@jhu.edu Abstract The evolution of social media users’ behav- ior over time complicates user-level compar- ison tasks such as verification, classification, clustering, and ranking. As a result, na¨ıve ap- proaches may fail to generalize to new users or even to future observations of previously known users. In this paper, we propose a novel procedure to learn a mapping from short episodes of user activity on social media to a vector space in which the distance between points captures the similarity of the corre- sponding users’ invariant features. We fit the model by optimizing a surrogate metric learn- ing objective over a large corpus of unlabeled social media content. Once learned, the map- ping may be applied to users not seen at train- ing time and enables efficient comparisons of users in the resulting vector space. We present a comprehensive evaluation to validate the benefits of the proposed approach using data from Reddit, Twitter, and Wikipedia. 1 Introduction Social media presents a number of challenges for characterizing user behavior, chief among them that the topics of discussion and their participants evolve over time. This makes it difficult to un- derstand and combat harmful behavior, such as election interference or radicalization (Thompson, 2011; Mihaylov and Nakov, 2016; Ferrara et al., 2016; Keller et al., 2017). This work focuses on the fundamental prob- lem of learning to compare social media users. We propose a procedure to learn embeddings of small samples of users’ online activity, which we call episodes. This procedure involves learning the embedding using a metric learning objective that causes episodes by the same author to map to nearby points. Through this embedding users may be efficiently compared using cosine similar- ity. This representation immediately enables sev- eral tasks: Verification. Determining if two episodes have the same author. Classification. nearest neighbors.Labeling authors via their k- Clustering. Grouping users via off-the-shelf methods like k-means or agglomerative clustering. Ranking and retrieval. Sorting episodes accord- ing to their distances to a given episode. The problem considered in this paper is most closely related to author attribution on social me- dia. However, prior work in this area has primarily focused on classifying an author as a member of a closed and typically small set of authors (Sta- matatos, 2009; Schwartz et al., 2013; Shrestha et al., 2017). In this paper, we are concerned with an open-world setting where we wish to character- ize an unbounded number of users, some observed at training time, some appearing only at test time. A further challenge is that the episodes being com- pared may be drawn from different time periods. With these challenges in mind, the primary contri- butions described in this …

Content truncated to 500 words