On the Accuracy of Hyper-local Geotagging of Social Media Content

Abstract

Social media users share billions of items per year, only a small fraction of which is geotagged. We present a datadriven approach for identifying non-geotagged content items that can be associated with a hyper-local geographic area by modeling the location distributions of n-grams that appear in the text. We explore the trade-off between accuracy and coverage of this method. Further, we explore differences across content received from multiple platforms and devices, and show, for example, that content shared via different sources and applications produces significantly different geographic distributions, and that it is preferred to model and predict location for items according to their source. Our findings show the potential and the bounds of a data-driven approach to assigning location data to short social media texts, and offer implications for all applications that use data-driven approaches to locate content.

Extracting fine‐grained location with temporal awareness in tweets: A two‐stage approach

“…Recently, Yuan, Cong, Ma, Sun, and Magnenat‐Thalmann () further investigated spatial, temporal, and topical aspects to model users' geographic activities. Flatow, Naaman, Xie, Volkovich, and Kanza () addressed the task of fine‐grained location estimation by identifying hyper‐local n ‐grams based on a collection of geo‐tagged tweets of a specific geographic area. Although these studies partially enable fine‐grained location estimation, the specific POI information may be lost due to ambiguity, and the temporal awareness is still unknown.…”

Location Inference of Social Media Posts at Hyper-Local Scale

“…al. [8] perform hyper-local geo-location as we do, but limit their analysis to the NYC area. This approach relies on the identification of geo-specfic n-grams.…”

“…About 750 km was reported as the mean distance error, considering tweets from the entire U.S. Some works [8,5] use n-grams extracted from the content.…”

“…Finally, feature selection offers our model a two-fold advantage. First, we can identify and geo-locate only those tweets that contain geographically relevant content unlike simply noting that tweets that lack contextual features is a limitation of the model [8]. Second, selection eliminates noisy features and boosts model performance.…”

“…Moreover, even when users (rarely) populate this location field it is usually too broad to be of any use or even fictitious [3]. Finally, the finest granularity at which several contemporary approaches infer location is that of a city [11,8,9,10,3], which cannot bring much value for most applications.…”