A Google Trends spatial clustering approach for a worldwide Twitter user geolocation

Abstract

User location data is valuable for diverse social media analytics. In this paper, we address the non-trivial task of estimating a worldwide city-level Twitter user location considering only historical tweets. We propose a purely unsupervised approach (no location data is used) that is based on a synthetic geographic sampling of Google Trends (GT) city-level frequencies of tweet nouns and three clustering algorithms. The approach was validated empirically by using a recently collected dataset, with 3,268 worldwide city-level locations of Twitter users, obtaining competitive results when compared with a state-of-the-art Word Distribution (WD) user location estimation method. The best overall results were achieved by the GT noun (GTN) DBSCAN (GTN-DB) method, which is computationally fast, and correctly predicts the ground truth locations of 15%, 23%, 39% and 58% of the users for tolerance distances of 250 km, 500 km, 1,000 km and 2,000 km.

Geo-Spatial Market Segmentation & Characterization Exploiting User Generated Text Through Transformers & Density-Based Clustering

“…As for practical implications, we recommend performing the spatial characterization as per city basis, this is mostly due to how the density based clusters vary according to the observed region and the computational power available. Moreover, in similar cases to Twitter, where none or only a small portion of messages carry the location context, incorporating other solutions to predict location context from messages lacking this information will significantly increase the observation surface [18]. Furthermore, practitioners might want to tune or extend upon our classification results, particularly for the classes with lower F1-scores.…”

“…We considered this to be of interest because marketing campaigns can be costly and may suffer from low user response compared to the original investment. Moreover, there might be other reasons to estimate the optimal location for a particular activity [17], such as opening a new store [6], [7], or estimating the interest for something in particular locations to validate assumptions [18]. Our work focuses on how to characterize geographic areas in relation to a selected set of product categories.…”

“…In particular, the lack of dynamic and geo tagged opinion data. However, in the case of Twitter, other researches have proposed approaches to infer the locations of messages when lacking this information [18], [53]. Furthermore, to the best of our knowledge, there were few public geo tagged free text datasets that we could exploit.…”

Approaches for the Clustering of Geographic Metadata and the Automatic Detection of Quasi-Spatial Dataset Series

“…The most used algorithms are distance-based solutions such as K-means, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), or Ordering Points To Identify the Clustering Structure (OPTICS), and probabilistic ones like Probabilistic Latent Semantic Indexing (PLSI) [20]. The work of Zola et al [40] is a good example of how some of these clustering techniques are currently used in the spatial data context to identify patterns in text collections. They estimate Twitter user location based on their tweets using Google Trends frequencies of tweet nouns and clustering to identify the most probable location.…”

Polarización en Twitter durante la crisis de la COVID-19: Caso Aislado y Periodista Digital

“…Diversos estudios estiman fiabilidades en función del lugar del tuit por encima del 90% en Reino Unido o Estados Unidos, 85,8% en España o 83,95% Filipinas, con una media mundial del 77,84% y 88,15% en Europa (van der Veen et al, 2015), o un error medio de localización de 256 kilómetros (Holbrook et al, 2016). Dichos porcentajes pueden ser mejorados mediante el uso de técnicas complementarias varias de las que hay diversa bibliografía, que no han sido aplicadas para este trabajo (Zola, Ragno y Cortez, 2020).…”