Interpreting Twitter User Geolocation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 853–859 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 853 Interpreting Twitter User Geolocation Ting Zhong1, Tianliang Wang1, Fan Zhou1†, Goce Trajcevski2, Kunpeng Zhang3, Yi Yang4 1University of Electronic Science and Technology of China, China 2Iowa State University, Ames, IA, USA 3University of Maryland, College Park, USA 4The Hong Kong University of Science and Technology, Hong Kong †Corresponding author: fan.zhou@uestc.edu.cn Abstract Identifying user geolocation in online social networks is an essential task in many location- based applications. Existing methods rely on the similarity of text and network struc- ture, however, they suffer from a lack of inter- pretability on the corresponding results, which is crucial for understanding model behavior. In this work, we adopt influence functions to interpret the behavior of GNN-based models by identifying the importance of training users when predicting the locations of the testing users. This methodology helps with providing meaningful explanations on prediction results. Furthermore, it also initiates an attempt to un- cover the so-called “black-box” GNN-based models by investigating the effect of individ- ual nodes. 1 Introduction Identifying geographic locations of users in online social networks (OSN) has become a key Internet service for many downstream applications, includ- ing location-based targeted advertising, emergency location identification, political election campaign, local event/place recommendation, natural disas- ter response, and remediation, etc. (Zheng et al., 2018). As such, the problem of user geolocation (UG) has received a great deal of research attention in the past decade (Han et al., 2012; Do et al., 2017; Miura et al., 2017; Rahimi et al., 2018; Bakerman et al., 2018). Earlier efforts (Amitay et al., 2004; Wing and Baldridge, 2011; Han et al., 2012; Roller et al., 2012; Ahmed et al., 2013; Han et al., 2014; Chong and Lim, 2017) mainly focused on extracting in- dicative information from user-posted contents. These approaches rely on informative words that can link users to their specific locations via vari- ous natural language processing techniques such as topic modeling and other statistical models. More recently, researchers aimed at incorporating multi-aspect information, especially the user men- tion/interaction network to boost the performance of geolocation identification (Rahimi et al., 2015; Do et al., 2017; Rahimi et al., 2017, 2018; Hamouni et al., 2019). For example, (Rahimi et al., 2018; Wu et al., 2019) employ Graph Convolutional Net- works (GCNs) (Kipf and Welling, 2017) or sim- plified GCN (Wu et al., 2019) to learn network structures for user geolocation. In addition, graph representation-based methods (Tang et al., 2015; Grover and Leskovec, 2016; Kipf and Welling, 2017; Hamilton et al., 2017; Qiu et al., 2018) have also been widely used for user geolocation (Do et al., 2017; Miura et al., 2017; Rahimi et al., 2018; Hamouni et al., 2019; Huang and Carley, 2019). However, the existing methods lack model trans- parency and fail to provide meaningful explana- tions regarding the model behavior and prediction results, which prevents them from safety-critical applications. For example, when locating an …

Content truncated to 500 words