Sage Journals: Discover world-class research

Abstract

Whenever someone posts an online review of a restaurant, museum, or barbershop, they also leave a trace of where they traveled. The author visualizes travel patterns in 11 North American metropolitan areas using geolocated review data. The data are based on approximately 7 million online Yelp.com reviews posted by 2 million reviewers between 2005 and 2020. First, the author demonstrates how individual travel patterns can be mapped using the review data and discusses the potential applications of such individual-level data. The author then turns to aggregate-level maps, creating establishment covisit networks in which two establishments are linked if multiple reviewers visit both. Maps of establishment covisits reveal various intriguing patterns related to consumption and geography, such as the connections between neighborhoods and the centralization and segregation within a metropolitan area. Establishment covisit maps can also inform researchers about the diffusion of ideas and practices, trends in crime, and gentrification.

Keywords

travel patterns consumption patterns urban sociology online reviews GIS mapping urbanization gentrification

The study of human mobility across urban neighborhoods is essential for understanding the dynamics of urban life and its broader societal impacts, such as gentrification, the diffusion of ideas, and the spread of crime. Historically, this field has seen diverse methodologies, ranging from direct observations and field studies (Jacobs 1961; Whyte 1980) to analog mapping techniques (Gould and White 1974), archival historical data (Braudel 1981), and manual traffic counts (Appleyard, Lynch, and Myer 1964; Levinson and Kumar 1997). The advent of Global Positioning System technology has shifted the focus toward electronic “breadcrumbs” left by individuals, with analyses of cell phone data and social media check-ins that shed light on various urban phenomena (Batty 2013; Buckee et al. 2020; Eagle, Macy, and Claxton 2010; Hui, Fader, and Bradlow 2009; Lu, Bengtsson, and Holme 2012).

In this visualization, I propose an alternative mapping approach using online review data, which may offer valuable insights into travel patterns in urban settings. Online reviews, as a data source, can augment existing sources such as cell phone data in studying movement, offering distinct strengths and weaknesses compared with cell phone data. The primary limitation is that online reviews represent a selected dataset: only a small proportion of people leave online reviews, and even active reviewers do not always review every establishment they visit. However, a significant advantage of these data is their public availability and inherent combination with consumption data.

This study analyzes approximately 7 million Yelp.com reviews posted across 11 North American metropolitan areas, revealing patterns in establishment covisits and their implications for urban and consumption sociology.

Data and Setting

I investigate establishment covisit patterns by examining publicly accessible online review data from Yelp.com, as provided in the 2022 Yelp Open Dataset, available at https://www.yelp.com/dataset. This dataset comprises 6,990,280 reviews from 1,987,897 users, covering businesses across different metropolitan areas, including Tucson, Arizona; Santa Barbara, California; Reno, Nevada; Boise, Idaho; Edmonton, Alberta, Canada; St. Louis, Missouri; Nashville, Tennessee; Indianapolis, Indiana; Philadelphia, Pennsylvania; Tampa, Florida; and New Orleans, Louisiana).¹ The dataset encompasses reviews of various establishments, such as post offices, bars, barbershops, and gyms; however, more than 80 percent of the reviews are related to restaurants. Each review includes several details, such as the date it was written, the identification number of the establishment being evaluated, and an anonymized ID for the person who wrote the review. In this visualization, I also use geolocation data for the establishments, provided in latitude-longitude format.

The dataset spans from November 2005 to January 2022. As the coronavirus disease 2019 pandemic lockdowns significantly altered dining patterns, which had not fully recovered by early 2022 (resulting in insufficient postpandemic observations in the dataset), I limited the observation window from November 2005 to February 2020.

I developed a Python script to calculate both individual and aggregate travel patterns, as well as their visualization; see Appendix B for the full code.

Visualizing Individual-Level Patterns

First, I visualize establishment covisit patterns on an individual level. Understanding individual travel patterns is informative for multiple domains in sociology, such as assessing the consumption habits of individuals, the breadth of their social networks, or their openness to different experiences. Individual-level maps can assist researchers in gaining a clearer understanding of reviewers’ travel and consumption patterns. These data could be used to address various questions, such as whether a person tends to stay local or covers a wide geographical area and whether they visit limited neighborhood types, such as highly affluent areas, or a more diverse range, including downtowns, suburbs, and neighborhoods of varying affluence. Additionally, these patterns could aid researchers in computing measures such as cosmopolitanism (Kovács and Carroll, 2023).

To visualize establishment covisit patterns, I combine the geolocation data of establishments visited by a reviewer and represent it as a directed network, where location A is connected to location B if a person visits location B following location A. Figure 1 illustrates a few such maps. As in this article I cannot include a separate map for each of the 2 million reviewers in the dataset, I picked two reviewers randomly. As the figure shows, the Indianapolis-based reviewer mostly visits establishments in the northern suburbs of Indianapolis and reviews a downtown destination only once. Contrast this with the reviewer from New Orleans, who typically frequents downtown locations and visits only three locations outside the downtown area. Information such as this can help characterize reviewer habits, especially when merged with neighborhood characteristics such as average income, neighborhood racial composition, or average education level.

Figure 1.

Mobility patterns of Yelp reviewers in Indianapolis and New Orleans: (a) individual-level Indianapolis, (b) individual-level New Orleans, (c) aggregate-level Indianapolis, and (d) aggregate-level New Orleans.

Visualizing Aggregate-Level Patterns

Although individual-level maps are illustrative, creating a map for each of the approximately 2 million reviewers in the dataset is not feasible. Instead, one can create aggregate-level maps to visualize typical travel patterns in a given metropolitan area. To achieve this, I create what are known as establishment covisit maps. Specifically, for each cosmopolitan area, I calculate the number of reviewers who visited both establishments in each pairwise combination within that area. (These networks are undirected and do not consider timing.) I then create a network visualization overlayed on the map in which two establishments are connected with a line if there are at least 10 reviewers who visited both establishments. Figures 1c and 1d show these maps for Indianapolis and New Orleans. Appendix A includes similar maps for the other nine metropolitan areas in the database. Appendix A also shows, for the case of the Philadelphia metropolitan area, how the choice of the threshold influences the visualization.² (This threshold can be easily modified in the Python script in Appendix B. The maps are also zoomable once the script has run.) These maps illustrate the traffic volume between establishments and neighborhoods. For example, the map of Indianapolis shows a hub-and-spoke travel structure with strong connections between downtown and the northern suburbs. The covisit patterns in New Orleans, on the other hand, show a more integrated travel pattern for most neighborhoods (less of a hub-and-spoke pattern); yet there are major neighborhoods in the city that are less frequented by Yelp reviewers (perhaps not surprisingly, these are the areas that were most heavily affected by Hurricane Katrina (https://www.nytimes.com/interactive/2015/08/25/us/mapping-katrina-and-aftermath.html).

Although the systematic analysis of these covisit travel patterns is beyond the scope of this article, research opportunities abound. For example, one could use such maps to document and possibly predict gentrification. Future research could also analyze whether ethnically diverse communities are more connected to the core or stay in the periphery (Eagle et al. 2010). One could explore if there are correlations between income levels, education levels, or demographics and people’s movements between businesses. Urban designers can use such maps to examine how urban design influences people’s movements. For example, do pedestrian-friendly areas or the availability of public transport encourage more movement between neighborhoods? For a more practical purpose, one could identify different customer segments on the basis of their movement patterns. For example, some customers may be authenticity seeking (Kovács, Carroll, and Lehman 2014) and thus connect highly authentic establishments and neighborhoods. Information such as this could provide insights on where to place advertising or how to optimize business hours.

Such maps could also inform studies of tourism. In related work, Hawelka et al. (2014) combined TripAdvisor reviews with transport data in Barcelona and London to model intracity movement flows of different traveler types. Business travelers moved between commerce districts, while other tourists congregated at attractions. Meanwhile, McKercher et al. (2012) inferred tourist mobility archetypes in Singapore from TripAdvisor trails, categorizing patterns such as “museum hoppers” and “resort lovers.” This mobility segmentation has applications for tourism planning.

Finally, I have shown maps for covisits only within metropolitan areas, but the same approach (and the script in Appendix B) could also be applied to study inter–metropolitan area covisits.

Conclusion, Limitations, and Avenues for Future Research

This article has demonstrated the possibility of using online review data to map spatial movement and consumption patterns in metropolitan areas. I provide methods for mapping spatial consumption patterns both at the individual and aggregate levels and suggest research questions for which such maps can be helpful in answering.

The use of online reviews as a data source is intended not to replace but to complement traditional sources such as cell phone data in the analysis of movement patterns. Mapping on the basis of online reviews has its own unique advantages and disadvantages compared with using cell phone data. Its main drawback is that it constitutes a selective dataset: only a limited number of individuals post online reviews, and those who frequently review do not necessarily cover every place they visit. On the other hand, a notable benefit of these data is that they are publicly accessible and intrinsically linked with information on consumer behavior.

Footnotes

Appendix: Additional Maps

Appendix B: Source Code

Here is the full Python 3 code to reproduce the figures.

## First, get the freely available data from https://www.yelp.com/dataset

import json

from collections import defaultdict

import pandas as pd

import folium

from math import radians, cos, sin, asin, sqrt

import math

import matplotlib.pyplot as plt

from folium.plugins import AntPath

import random

from folium.vector_layers import CircleMarker

input_file1 = "yelp_academic_dataset_review.json"

input_file2 = "yelp_academic_dataset_business.json"

output_file = "dyad.txt"

# Function to calculate distance between two points using Haversine formula

def haversine(lon1, lat1, lon2, lat2):

lon1, lat1, lon2, lat2 = map(radians, [lon1, lat1, lon2, lat2])

dlon = lon2 - lon1

dlat = lat2 - lat1

a = sin(dlat/2)**2 + cos(lat1) * cos(lat2) * sin(dlon/2)**2

c = 2 * asin(sqrt(a))

# Radius of earth in miles. Use 6371 for kilometers

r = 3956

return c * r

users_businesses = defaultdict(set)

# Load the review JSON file

with open(input_file1, 'r') as infile1:

lines_read = 0

for line in infile1:

data = json.loads(line)

user_id = data.get('user_id', '')

business_id = data.get('business_id', '')

users_businesses[user_id].add(business_id)

lines_read += 1

# If 20000 lines have been read, break the loop

#if lines_read >= 1000000:

# break

# Create a dictionary to store the latitude and longitude for each business

business_locations = {}

# Load the business JSON file

with open(input_file2, 'r') as infile2:

for line in infile2:

data = json.loads(line)

business_id = data.get('business_id', '')

latitude = data.get('latitude', '')

longitude = data.get('longitude', '')

business_locations[business_id] = (latitude, longitude)

# Create a dictionary where the keys are pairs of businesses and the values are the counts of users who visited both businesses

business_pairs = defaultdict(int)

# For each user, consider each pair of businesses the user visited

for businesses in users_businesses.values():

for business1 in businesses:

for business2 in businesses:

if business1 < business2: # to avoid duplicating pairs

business_pairs[(business1, business2)] += 1

# Store the dyads' locations in an array# Write the pairs of businesses with at least 5 common users to the output file

dyads = []

with open(output_file, 'w') as outfile:

outfile.write("business1\tbusiness1_latitude\tbusiness1_longitude\tbusiness2\tbusiness2_latitude\tbusiness2_longitude\tcount\n")

for (business1, business2), count in business_pairs.items():

if count >= 10:

# Get the latitude and longitude for each business

latitude1, longitude1 = business_locations.get(business1, ('', ''))

latitude2, longitude2 = business_locations.get(business2, ('', ''))

# calculate distance between two businesses

# to select a specific latitude and longitude, use the following code

if ((int(latitude1) == 39 or int(latitude1) == 40) and int(longitude1) == -75) :

#if latitude1 != '' and longitude1 != '' and latitude2 != '' and longitude2 != '':

distance = haversine(float(longitude1), float(latitude1), float(longitude2), float(latitude2))

if distance <= 100: # only consider businesses within 100 miles

outfile.write(f"{business1}\t{latitude1}\t{longitude1}\t{business2}\t{latitude2}\t{longitude2}\t{count}\n")

dyads.append(((latitude1, longitude1), (latitude2, longitude2),count))

# Create the map centered at the first pair of coordinates

m = folium.Map(location=[float(dyads[0][0][0]), float(dyads[0][0][1])], zoom_start=10)

# Iterate over the pairs of coordinates and create a line for each

for dyad in dyads:

folium.PolyLine([dyad[0], dyad[1]], color="blue', opacity=math.log(dyad[2])/100).add_to(m)

# Save the map to an HTML file –> this file can be opened in a browser and one can move around the map and zoom in and out

m.save('map25_zoom10_philadelphia.html')

# Filter reviewers who have between 10 and 15 reviews

eligible_users = [user for user, businesses in users_businesses.items() if 15 <= len(businesses) <= 25]

selected_user = random.choice(eligible_users)

# Extract establishments visited by this reviewer

visited_businesses = users_businesses[selected_user]

# Extract locations of these establishments

visited_locations = [business_locations[business_id] for business_id in visited_businesses if business_id in business_locations]

# Create a new map centered at the location of the first establishment visited by the reviewer

if visited_locations:

reviewer_map = folium.Map(location=[float(visited_locations[0][0]), float(visited_locations[0][1])], zoom_start=10)

# Plot these establishments on the map

for latitude, longitude in visited_locations:

folium.Marker([float(latitude), float(longitude)]).add_to(reviewer_map)

# Save this map to a separate HTML file

reviewer_map.save('reviewer_map.html')

def draw_map_for_user(user_id, business_locations, users_businesses):

visited_businesses = users_businesses[user_id]

visited_locations = [business_locations[business_id] for business_id in visited_businesses if business_id in business_locations]

if visited_locations:

reviewer_map = folium.Map(location=[float(visited_locations[0][0]), float(visited_locations[0][1])], zoom_start=10)

# Connect the establishments with blue arrows

for i in range(len(visited_locations) - 1):

latlon1 = (float(visited_locations[i][0]), float(visited_locations[i][1]))

latlon2 = (float(visited_locations[i + 1][0]), float(visited_locations[i + 1][1]))

folium.PolyLine([latlon1, latlon2], color="blue').add_to(reviewer_map)

reviewer_map.save(f'reviewer_map_{user_id}.html')

# Filter reviewers who have between 10 and 15 reviews

eligible_users = [user for user, businesses in users_businesses.items() if 10 <= len(businesses) <= 15]

# Randomly select 10 such reviewers without replacement

selected_users = random.sample(eligible_users, 10)

# Iterate over the selected users, print their user_id and draw the maps

for user_id in selected_users:

print(user_id)

draw_map_for_user(user_id, business_locations, users_businesses)

ORCID iD

Balázs Kovács

Notes

Author Biography

Balázs Kovács is a professor of management in the Yale School of Management and the Yale Department of Sociology (courtesy). He studies various topics in organization theory and computational social science, including social networks, online reviews, authenticity, learning, diffusion, identity, text analysis, and status. His current work investigates the effects of category spanning and innovation in technological and cultural domains.

References

Appleyard

Donald

Lynch

Kevin

Myer

John R.

. 1964. The View from the Road. Cambridge, MA: MIT Press.

Batty

Michael

. 2013. “Big Data, Smart Cities and City Planning.” Dialogues in Human Geography 3(3):274–79.

Braudel

Fernand

. 1981. The Structures of Everyday Life: Civilization and Capitalism 15th–18th Century, Vol. 1. Berkeley: University of California Press.

Buckee

Caroline O.

Balsari

Satchit

Onnela

Jukka-Pekka

Santillana

Mauricio

Grad

Yonatan H.

. 2020. “Aggregated Mobility Data Could Help Fight COVID-19.” Science 368(6487):145–46.

Eagle

Nathan

Macy

Michael

Claxton

Rob

. 2010. “Network Diversity and Economic Development.” Science 328(5981):1029–31.

Gould

Peter R.

White

Rodney

. 1974. Mental Maps. New York: Penguin.

Hawelka

Bartosz

Sitko

Izabela

Beinat

Euro

Sobolevsky

Stanislav

Kazakopoulos

Pavlos

Ratti

Carlo

. 2014. “Geo-located Twitter as Proxy for Global Mobility Patterns.” Cartography and Geographic Information Science 41(3):260–71.

Hui

Sam K.

Fader

Peter S.

Bradlow

Eric T.

. 2009. “Path Data in Marketing: An Integrative Framework and Prospectus for Model Building.” Marketing Science 28(2):320–35.

Jacobs

Jane

. 1961. The Death and Life of Great American Cities. New York: Random House.

10.

Kovács

Balázs

Carroll

Glenn R.

. 2023. “Distinguishing Between Cosmopolitans and Omnivores in Organizational Audiences.” Academy of Management Discoveries 9(4):549–77.

11.

Kovács

Balázs

Carroll

Glenn R.

Lehman

David W.

. 2014. “Authenticity and Consumer Value Ratings: Empirical Tests from the Restaurant Domain.” Organization Science 25(2):458–78.

12.

Levinson

David M.

Kumar

Ajay

. 1997. “Density and the Journey to Work.” Growth and Change 28(2):147–72.

13.

Xin

Bengtsson

Linus

Holme

Petter

. 2012. “Predictability of Population Displacement After the 2010 Haiti Earthquake.” Proceedings of the National Academy of Sciences 109(29):11576–81.

14.

McKercher

Bob

Shoval

Noam

Eric

Birenboim

Amos

. 2012. “First and Repeat Visitor Behaviour: GPS Tracking and GIS Analysis in Hong Kong.” Tourism Geographies 14(1):147–61.

15.

Whyte

William H

. 1980. The Social Life of Small Urban Spaces. Washington, D.C.: The Conservation Foundation.