Addressing the location A/B problem on Twitter: the next generation location inference research

Abstract

Often, global and regional topics on Twitter across multiple thematic areas, such as disasters, politics, protests, entertainment, epidemics, literature, travel, culture, weather, etc., witness an unprecedented level of exchange of conversations. An issue with those conversations is that a user can be at location A and participate in a public discourse specific to location B, which we refer to as the Location A/B problem. Location profiling of users solely based on locations mentioned in their tweets leads to ineffective location-based recommendations. The problem is deemed solved if location candidates could be categorized as either origin locations (Location As) or non-origin locations (Location Bs); however, real-world tweets are much more complex, and currently, no public datasets are available for training such classifiers. To the best of our knowledge, this study yields the first steps in addressing the Location A/B problem on Twitter. We propose a theoretical framework that utilizes the existing literature on location inference to categorize location candidates as either origin locations or non-origin locations. We envision that: (i) the framework provides the grounds for designing models that aim to solve the Location A/B problem, and (ii) the location profiling of users based on origin locations leads to improved geotargeted recommendations.