Geotagged tweets trace how pandemic reshaped New York City neighborhood chatter

The study, published in the journal Computational Urban Science, classifies tweets into five everyday urban domains—transportation, retail activity, cultural and social activity, healthcare, and work or remote work—and tracks how each one flared, collapsed, and reorganized in space and time.

The research was carried out by Marjorie Mattes, Dorian Arifi, and Bernd Resch, who mined a dataset of public, geotagged Twitter posts collected through the platform’s Streaming and REST APIs by the University of Salzburg. Every record included the tweet text, a timestamp, and precise GPS coordinates, allowing the team to work at the level of ZIP Code Tabulation Areas across the five boroughs of Brooklyn, the Bronx, Manhattan, Queens, and Staten Island. After filtering out non-English content, which removed about 21 percent of the corpus, and excluding institutional or automated accounts using a keyword-based heuristic, the researchers narrowed roughly 15 million tweets down to 2,281,953 category-relevant posts within the city—about 15 percent of the original dataset.

Classifying millions of short, noisy social media messages into meaningful categories is a notoriously difficult problem. Keyword searches alone retrieve thematically relevant content but stumble when categories overlap semantically, while unsupervised topic models applied directly to raw Twitter streams can produce incoherent clusters. The team’s hybrid workflow addressed this by first building keyword lists from Wikipedia: they scraped seed articles for each functional category, embedded the resulting corpus with the Sentence-BERT model all-MiniLM-L6-v2, clustered the articles with K-means, and extracted high-scoring terms with TF-IDF before manual curation. Tweets containing any of these keywords then entered BERTopic, an embedding-based topic modeling pipeline that uses UMAP for dimensionality reduction and HDBSCAN for density-based clustering, run separately on each year of data with GPU acceleration.

The pipeline was deliberately tuned for precision over completeness. In a manual evaluation based on 200 randomly sampled tweets per category, every category achieved precision scores above 90 percent, meaning that tweets labeled as belonging to a domain very often genuinely belonged there. The cost came in recall: of 948 manually relevant tweets in a labeled validation sample, only 209 survived the full workflow, an overall recall of 22.0 percent, with a strict category-level accuracy of 17.7 percent. BERTopic itself assigned 39.28 percent of the filtered tweets to distinct topics, leaving 60.72 percent as outliers. The authors frame the resulting category counts as conservative indicators of clearly identifiable activity rather than complete representations of urban behavior.

Within those constraints, the temporal patterns were striking. Cultural and social activity dominated the conversation, accounting for between 21.99 and 36.71 percent of category-related tweets and peaking in 2021. Retail activity fell steadily, from a high of 7.86 percent of category-related tweets in 2019 to 5.17 percent in 2022. Transportation moved in the opposite direction, climbing from just 2.90 percent in 2018 to 14.18 percent in 2022. Healthcare peaked at 2.5 percent of the category-related share in 2020, otherwise ranging between 0.48 and 1.23 percent, while work and remote work rose from a low of 0.67 percent in 2019 to 2.08 percent in 2022. Category-relevant tweeting overall peaked in 2019 and declined in subsequent years, mirroring a broader drop in geolocated tweeting across the metropolitan area.

To separate genuine disruption from ordinary fluctuation, the researchers normalized tweet frequencies by the total number of tweets in each ZIP code and measured each area’s deviation from its long-term 2018-to-2022 mean, following an approach used in prior urban analytics work. They then tested whether those deviations were spatially structured using Global and Local Moran’s I statistics with a fixed-distance band. The distinction proved crucial. In transportation, significant spatial autocorrelation appeared in 2018, vanished during 2019 through 2021, and returned highly significant by 2022, when high-high clusters concentrated along subway and railway corridors in Manhattan and Brooklyn while Queens and the Bronx showed extensive low-low clusters. A persistent high-low outlier sat around Jamaica Station in central Queens, which had shown above-average activity as early as 2020 even as JFK Airport tracked its long-term average.

Cultural and social activity told a different story. Before the pandemic, tweeting about parks, recreation, and social life was spread broadly across all boroughs, roughly matching the citywide distribution of cultural facilities. In 2020, an above-average corridor emerged between Brooklyn and Queens, partly bordering the COVID-19 restriction zones introduced in autumn 2020, while Staten Island areas with large parklands and the eastern shoreline of the Bronx also showed elevated shares. Statistically significant local clusters appeared mainly in Staten Island, which the authors link to the borough’s extensive parks and lower-density structure—consistent with earlier research finding that New York’s nature areas saw stable or increased visitation during the pandemic. Notably, the visually obvious Brooklyn-Queens corridor did not register as a statistically significant cluster, a reminder that striking map patterns and robust spatial statistics are not the same thing.

Retail activity proved the most fragmented domain. Global Moran’s I detected significant clustering only in 2019, and even then Local Moran’s I found no coherent local clusters, including in the visually intensified area of northeastern Queens. Localized increases near commercial centers such as King’s Plaza and Gateway Center in southeastern Brooklyn in 2022 likewise lacked statistical support and should be read as visual tendencies rather than established neighborhood patterns. The authors suggest this heterogeneity reflects the mixed nature of retail tweets, which can range from consumer reviews and complaints to business promotions and advertisements, each following different spatial dynamics. Healthcare and work or remote work, meanwhile, produced signals too sparse and spatially patchy to support robust interpretation at all, despite their obvious conceptual relevance to the pandemic—a finding the authors highlight as evidence that thematic importance does not guarantee analytical usefulness in geotagged social media data.

The team emphasizes that the workflow was designed as a retrospective analytical tool rather than a monitoring system, and that its spatial validation rested on correspondence with urban features—transit lines, parks, commercial centers—rather than direct comparison with official statistics. Spatial correspondence alone cannot establish causality or confirm that tweet activity reflects actual behavioral change; the transportation clustering in 2022, for instance, indicates that transit infrastructure became more visible in geotagged discourse, not that ridership recovered. Other limitations compound the caution: geotagged tweets represent a digitally active minority, the analysis was English-only in a famously multilingual city, low-volume ZIP codes can produce large relative deviations from small absolute changes, and residual automated or organizational accounts may have influenced some patterns. The authors propose future work linking tweet patterns to station-level passenger counts, retail footfall, and park visitation data, extending the framework to other cities and languages, and running it at weekly or monthly resolution to capture responses to policy changes and infection waves in near real time.

Subject of Research: Spatiotemporal analysis of geotagged tweets classifying New York City urban activity across the COVID-19 pandemic

Article Title: Understanding pre- and post-COVID urban neighborhood dynamics: a study analyzing georeferenced tweets for New York City

Article References: Mattes, M., Arifi, D., & Resch, B. (2026). Understanding pre- and post-COVID urban neighborhood dynamics: a study analyzing georeferenced tweets for New York City. Computational Urban Science, 6(1), Article 58. https://doi.org/10.1007/s43762-026-00293-2

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00293-2

Keywords: geotagged tweets, New York City, COVID-19 pandemic, BERTopic, urban analytics, spatial autocorrelation, transportation, cultural activity, retail activity, remote work, social media data, ZIP codes