Hyperlocal home location identification of Twitter profiles

Knowledge of user's location provides valuable information that can be used to build region-specific models (e.g. language used in a particular region and map-based visualisations of social media posts). Determining a user's home location presents a challenge. Current approaches make use of geo-located tweets or textual cues but are often only able to predict location to a coarse level of granularity (e.g. city level), while many applications require finer-grained (hyperlocal) predictions. A novel approach for hyperlocal home location identification, based on clustering of geo-located tweets, is presented. A goldstandard data set for home location identification is developed by making use of indicative phrases in geo-located tweets. We find that the cluster-based approaches outperform current techniques for hyperlocal location prediction.