Popularity Trajectories and Mixing Patterns of Urban Venues in Two European Cities

INTRODUCTION

• Understanding how venues gain, maintain, or lose popularity matters because venue successes and failures contribute to the health of urban spaces and neighborhood vibrancy.The introduction connects this question to business owners, urban planners, and nearby residents.Source: INTRODUCTION · p. 1
• Venue popularity is defined as a community’s relative response to a venue compared with competitors, measured through longitudinal signals such as Foursquare likes and check-ins.The paper motivates trajectory prediction as useful for owners’ strategic planning and timely correction to avert failure.Source: INTRODUCTION · pp. 1–2
• Potential influences include venue quality, pricing and branding; reviews; clustered similar businesses; and neighborhood income, density, and existing venues.The paper distinguishes clustering-related agglomeration from neighborhood effects.Source: INTRODUCTION · p. 2
• The study examines whether urban environments predict popularity trajectories, using Foursquare data for more than 10^4 venues in London and Barcelona followed for up to 5.5 years.It compares popularity time series with Dynamic Time Warping, groups similar series into six shapes, and assesses feature importance with permutation importance.Source: INTRODUCTION · p. 2
• Neighborhood socioeconomic status and density do not predict trajectories, while nearby venues’ trajectories mirror one another in a pattern consistent with local performance spillover.Same-shape venues appear together more often than chance within a few hundred meters; this pattern weakens at larger geographic coverage.Source: INTRODUCTION · p. 2
• Some trajectories become distinguishable well before the observation period ends, while others are recognizable later; the study also reads contrasting districts in London and Barcelona against their histories and local contexts.The districts are described as rising in London and falling in Barcelona.Source: INTRODUCTION · p. 2

Dataset and popularity measure

• For London and Barcelona, venues existing at least 130 weeks with at least 13 likes (5,923 and 4,718, respectively) have like counts aggregated into six-month bins and normalized by total popularity across all venues in each period.Source: Dataset and popularity measure · p. 3

Grouping trajectories into recurring shapes

• Dynamic Time Warping aligns venue series before KMeans groups them into six recurring shapes, with the gap statistic selecting the cluster count.Each series is represented as 11 six-month periods for fixed-size model inputs.Source: Grouping trajectories into recurring… · p. 3, Dataset and popularity measure · p. 3

Urban features

• Urban context is represented by area socioeconomic status and population density, with socioeconomic indices rescaled for cross-city comparison.The Barcelona socioeconomic score is sign-reversed so that higher values indicate less wealthy areas.Source: Urban features · pp. 3–4

Venue features

• Venue features measure nearby same-category competition and same-shape assortativity within city-specific radius circles.Review sentiment is also included for early classification, but is not reported in the main results.Source: Venue features · pp. 4–6
• A mixing matrix compares observed trajectory-shape co-occurrences with random-mixing expectations, while a spatial null test reshuffles labels at fixed coordinates.Circles with fewer than five venues are excluded; the test uses four radii per city without correcting for multiple tests.Source: Venue features · pp. 5–6

Classification and Forecast Models

• Early classifiers predict trajectory shape from truncated series, with and without urban and venue features; a GRU encoder-decoder forecasts normalized popularity values.Classification tests varying observed time lengths; forecasting compares joint training across venues with separate models for each trajectory cluster.Source: Classification and Forecast Models · pp. 6–7, Figure 1 · p. 6

Qualitative case study

• Hackney and Ciutat Vella were selected as contrasting cases with similar deprivation and density but sharply divergent venue trajectories, for interpretation rather than causal testing.The case studies use district histories and social theories to interpret street-level dynamics.Source: Qualitative case study · p. 7

Venue popularity follows six recurring shapes across cities

• Six recurring shapes appear in both cities: Fall, Rise, Rise & Fall, Steady, Ephemeral, and Underdogs; no recurring Fall & Rise shape emerged.Steady is the largest cluster (3,165 venues) and Underdogs the smallest (493).Source: Venue popularity follows six recurring… · p. 7
• Urban features alone are not predictive, while venue features reach F1 scores of 0.81 on the binary task and 0.71 on the six-class task.Across non-linear models and both cities, cluster assortativity is the strongest predictor; the authors suggest geographically clustered venues may rise or decline together.Source: Venue popularity follows six recurring… · pp. 7–8

How early a venue’s performance is visible depends on when its signature appears, not how complex the shape is

• At full history, XGBoost reaches macro F1 0.83; Ephemeral is easiest to classify (mean F1 = 0.86), while Rise & Fall is hardest (mean F1 = 0.54).Underdogs follow at mean F1 = 0.56.Source: How early a venue’s performance is… · p. 8
• Classification becomes possible at different times because shape signatures emerge at different points: Fall is apparent early, but Rise and Rise & Fall remain indistinguishable until later.Rise F1 stays flat until k = 9 and increases sharply at k = 10; Underdogs F1 starts increasing when k ≥8. Added urban and venue features generally provide little or inconsistent improvement, apart from a small Underdogs gain at k = 2–3.Source: How early a venue’s performance is… · p. 9
• Forecasting error falls from 0.224 at a 30-month lead (k = 5) to 0.099 at a six-month lead (k = 10); per-cluster training modestly outperforms joint training (mean MSE 0.171 vs. 0.190).Steady is most forecastable and Underdogs least forecastable under joint training; six-month forecasts remain a short practical planning window.Source: How early a venue’s performance is… · pp. 9–10, Table 1 · p. 10

Trajectory types mix at the block scale, not the neighbourhood

• At 300 m, most trajectory types occur near their own kind more often than chance, but Underdogs and Ephemeral show contrasting city-specific patterns.The remaining Fall, Steady, and Rise & Fall types strongly co-occur with one another in London.Source: Trajectory types mix at the block… · pp. 11–12

A comparative case study of venue clusters in Hackney and Ciutat Vella

• The case studies contrast tightly clustered Rise venues in Hackney with clustered Fall venues in Ciutat Vella, linking street-local popularity patterns to distinct gentrification histories.The case study cannot quantify policy or property speculation’s contribution, and cautions against asserting simple spillover mechanisms without historical context.Source: A comparative case study of venue… · pp. 12–14

DISCUSSION

• Venue success includes flourishing, stagnating, declining, recovering, and steadily climbing paths that binary survival analysis cannot distinguish.The study follows popularity over time to examine these trajectories and their surroundings.Source: DISCUSSION · p. 14
• Socioeconomic status and density carry little predictive power, while immediate neighbours’ trajectories predict a venue’s trajectory at street rather than neighbourhood or ward scale.The authors suggest that census data aggregated at larger scales may average over the scale at which the effect is observable.Source: DISCUSSION · p. 14
• The analysis is limited to Foursquare like counts in two European cities, with socioeconomic status measured differently and at different spatial scales across them.The authors call for cross-continent research to assess whether the trajectory shapes generalize.Source: DISCUSSION · p. 14
• The study is venue-type agnostic and recommends testing whether trajectories also appear using visitation or comment counts.The authors also note that alternative popularity signals may inform decision-makers more accurately and reliably.Source: DISCUSSION · p. 15
• The authors conclude that vibrancy is an emergent local process and suggest tailoring commerce-monitoring or support tools to immediate catchments rather than wards or boroughs.They describe combining behavioural trajectories with spatial and historical context as a route to studying urban vitality across cities and signals.Source: DISCUSSION · p. 15

DECLARATIONS

• The paper defines abbreviations including DTW, POI, IMD, MSE, GRU, ETSC, LLM, KDE, IQR, MAD, and XGBoost.Source: DECLARATIONS · p. 15

A DATA DESCRIPTION

• The like-count rank-frequency distribution has fitted exponents of approximately 2.34 and 2.50, but is not distinguishable from a lognormal distribution by the Vuong closeness test.The authors therefore describe it as a heavy-detailed distribution rather than a power law.Source: Figure S1 · p. 19
• Venue lifespan is measured as the number of weeks between the earliest and latest like made for each venue.Source: Figure S2 · p. 19
• Assortativity-coefficient distributions for London and Barcelona are centred around zero.Source: Figure S3 · p. 20
• The category figure uses each venue’s first appearing tag as its primary category, with pubs and Mediterranean food venues given as city-specific popular examples.Source: Figure S4 · p. 20
• Sentiment is measured with a multilingual DistilBERT model, rescaled to -2 to 2, aggregated every six months, and normalized to 11 time points.Missing values are imputed by the vector mean; length normalization uses linear interpolation and extrapolation.Source: Figure S5 · p. 21

B DESCRIPTION AND DATA OF CLUSTER TYPES

• DTW clustering recovered six popularity-trajectory types common to London and Barcelona.Source: Table 3 · p. 21
• Because Steady and Rise & Fall trajectories look similar, the authors fitted each to a quadratic form and compared average curvature.Source: Figure S6 · p. 22
• Rise & Fall trajectories had greater average curvature than Steady trajectories (average a = 0.021 vs. 0.016; Mann-Whitney-U U = 3329523.0, p < .001).The comparison used a two-sided Mann-Whitney-U test.Source: Figure S6 · p. 22

C.1 Baseline trajectory type classification performance

• Table 4 presents per-class F1-scores across all models, highlighting the easiest, hardest, and best-performing results.Source: Table 4 · p. 22

C.2 Trajectory type classification performance with urban and venue features

• Table 5 reports binary classification performance by city, feature set, and model using 5-fold cross-validation.It gives mean ± standard deviation and marks the best model in each feature-set block.Source: Table 5 · p. 23

• Table 6 reports multiclass classification performance by city, feature set, and model using 5-fold cross-validation.Results are shown as mean ± standard deviation, with the best model in each feature-set block marked in bold.Source: Table 6 · p. 24

• For k ≤9, Baseline_venue, Baseline_static, and All consistently outperform Baseline, Baseline_Sentiment, and Baseline_urban; all ablations rise sharply near k=10.The increase is consistent with the sequence becoming sufficiently informative once most of it has been observed.Source: Figure S7 · p. 24
• Table 7 compares London and Barcelona F1-scores across feature sets and models for binary and multiclass tasks.It defines ∆F1 as F1(London) − F1(Barcelona).Source: Table 7 · p. 25

• Table 8 reports mean ± standard deviation of permutation importance for predicting venue popularity trend with Logistic Regression, Random Forest, and XGBoost.Bold cells identify the most important feature for each model.Source: Table 8 · p. 26
• Table 9 compares the spatial null model for ˆα in London and Barcelona; an asterisk marks a significant difference between observed and null variance.Source: Table 9 · p. 26