Mobility Pattern Analysis during Russia–Ukraine War Using Twitter Location Data

Abstract

This paper aims to use location-based social media data to infer the impact of the Russia–Ukraine war on human mobility. We examine the impact of the Russia–Ukraine war on changes in human mobility in terms of the spatial range of check-in locations using social media location data. Specifically, we collect users’ check-in location data from Twitter and analyze the average gyration of check-ins from a region across the timeline of major events associated with the war. Change-point detection is performed on these time-series check-ins to identify the timeline of abrupt changes, which are shown to be consistent with the timing of a series of sanctions and policies. We find that war-related events may contribute secondary impacts (e.g., the surge in gas prices) to users’ travel patterns. The impact of the Russia–Ukraine war on users’ travel patterns can differ based on their own scope. Our case study demonstrates that users’ gyration in Warsaw, Paris, and Berlin experienced a decrease of over 50% during periods of gas price surges. These changes in users’ gyration patterns were particularly noticeable in neighboring countries like Poland compared to the other three countries. The findings of this study can assist policymakers, regulators, and urban planners to evaluate the impact of the war and to be adaptable to city planning after the war.

1. Introduction

This paper aims to use location-based social media data to infer the impact of the Russia–Ukraine war, which began in February 2022 and is still on-going, on human mobility. On 21 February 2022, Russian troops were sent into the Donetsk and Luhansk oblasts (administrative regions) in Ukraine after weeks of extreme tensions [1]. The order 14,065 (i.e., Blocking Property of Certain Persons and Prohibiting Certain Transactions) was then executed by the U.S. government [2]. After 23 February 2022, the European Union (EU) agreed on a range of packages of sanctions against Russia, covering finance, energy, transport, and technology sectors [3]. Undoubtedly, the on-going Russia–Ukraine war has triggered a series of global events with profound impacts on energy securities, economic relations, and financial markets.

The goal of this paper is to examine the impact of the Russia–Ukraine war on changes in human mobility in terms of the spatial range of check-ins in social media, which is used potentially as a surrogate of the regional population size and human mobility range. It has been 15 months since the war began, which has affected many countries, regions, and millions of populations [4]. Conducting travel pattern analysis over wide areas for a long period could be challenging in such a turbulent time. Therefore, we believe social media geo-tagged data offers an unintrusive, low-cost, and rapid way of evaluating human mobility patterns, when data collection with other methods, such as mobile phone or GPS, becomes infeasible, inaccessible, and potentially unethical.

The development of social media and the rapid advancement of technologies like big data, AI, and IoT have fundamentally changed our way of accessing information and knowledge. Social media data have consequently empowered our understanding of social activity and human mobility. To examine the war’s impact on human mobility, we utilize spatial–temporal check-ins from social media activities to approximate the mobile traces of individual users and estimate their travel patterns. We focus on the EU regions and analyze the potential impact of a series of sanctions and policies on human mobility. Specifically, we collect users’ check-in location data from Twitter and analyze the average gyration of check-ins from a region across the timeline of major events associated with the war. Change-point detection is performed to identify the timeline of major events solely from the Twitter data. The findings of this study can assist policymakers, regulators, and urban planners to evaluate the impact of the war on regional planning to be adaptable to city planning and economy after the war.

This research studies changes in mobility behaviors after the war by utilizing social media check-in data. The contribution of this work lies in the event-driven analysis of the impact of the Russia–Ukraine war on users’ travel patterns: (1) We examine the trends in check-ins, gyration, and distance across multiple cities, including Warsaw, Paris, and Berlin; (2) We explore the impact of global and local events on users’ gyration and distance patterns, and the secondary impacts of the war.

The rest of the paper is organized as follows: Section 2 describes the social media check-in data we collected. Section 3 introduces the methodology for analyzing gyration and distance patterns based on the check-in data. Section 4 presents a case study examining the gyration and distance performance in three cities. Section 5 concludes.

Related Work

Most studies focus on understanding the impact of the Russia–Ukraine war on financial markets [5], economy [6], and politics [7]. Some studies have examined the trend of aggregate travel demand following the Russia–Ukraine war. Ref. [8] investigates the reduced travel demand to Europe in the tourism industry. [9] leverages actual travel time between origins and destinations in Ukraine to infer changes in travel demand. This paper utilizes social media check-in data to study users’ travel patterns. In recent years, there has been a growing trend of using social media data in travel pattern studies [10,11,12,13,14,15,16]. This includes applications in crowd sensing and routing [17], activity pattern classification [18], activity location, and pattern inference [19]. There are also other studies that utilize social media data to understand the public’s attitude toward the war [20,21]. By leveraging social media data, researchers can complement traditional survey data or other open data sources and gain a more comprehensive understanding of the impact of the war.

2. Data Description

2.1. Twitter Location Data

We first introduce social media check-in data collected from Twitter. Geo-tagged posts in social media platforms offer users’ spatial–temporal (S-T) data through multi-day check-in records. To collect the check-in data, we initially utilized the Twitter API to retrieve tweets posted by individual users. Each tweet includes the user ID, user name, tweet ID, and timestamp. Subsequently, we employed the Tweepy package in Python 3 to gather each user’s historical tweets from 2021 and 2022. These historical tweets encompass the user ID, user name, timestamp, and geo-location (latitude and longitude), provided there are no privacy restrictions.

We filtered out historical tweets that lacked geo-locations, resulting in the remaining spatial–temporal check-ins used in this research. Each check-in record includes information about when, where, and by whom the check-in was made. We collected millions of check-in records from various major cities in Europe, including Paris, Warsaw, and Berlin. The number of check-ins and users are summarized in

Table 1. Daily check-in.

City	Mean	Std	Min	Median	Max
Warsaw 2022	1403.27	285.55	680	1392	2335
Warsaw 2021	1003.25	156.98	597	994.50	1572
Paris 2022	1821.21	446.53	894	1824	3421
Paris 2021	2362.91	357.83	1393	2390	3744
Berlin 2022	1733.55	289.71	1013	1728	2865
Berlin 2021	1191.55	159.15	806	1186	2017

and

Table 2. Daily user.

City	Mean	Std	Min	Median	Max
Warsaw 2022	303.14	75.98	141	297	595
Warsaw 2021	208.21	30.65	140	201.50	299
Paris 2022	301.20	73.51	166	290	493
Paris 2021	438.64	61.12	262	446	561
Berlin 2022	310.83	46.74	218	306	453
Berlin 2021	250.75	27.39	182	253	322

, respectively. The number of check-ins refers to the total count of daily check-ins, meaning it represents the aggregate number of check-ins that occur on each individual day. To calculate the average number of check-ins, we divided the total number of check-ins by the number of users for a given day. Figure 1 visualizes the number of check-ins across three cities over a 24 h period. Notably, activity peaks are evident around 8 a.m. and 8 p.m., while the trough is observed between 2 a.m. and 3 a.m. in the early morning. Our data align closely with the typical diurnal activity patterns observed on the social media platform. To better understand social media check-in data, we visualized the distribution of users in a log–log plot (see Figure 2). The dashed lines denote the fitting curves using power law functions. Approximately 80% of users have fewer than 50 spatial–temporal (S-T) records, indicating sporadic check-in behavior. Around 11.3% of users have between 50 and 200 S-T records, suggesting more regular engagement. Additionally, 4.9% of users exhibit a moderate level of activity with 200 to 500 S-T records, while 3.8% of users demonstrate a high level of engagement with over 500 S-T records. The orange dashed line indicates that the relative change of the number of users in Berlin, with respect to a one-unit increase in check-in numbers, is higher than that in other cities.

2.2. War-Related Events

To gain insights into the changes observed in users’ travel patterns, we employ an event-driven analysis in this work. The analysis allows us to explore the relationship between users’ gyration patterns and events that have the potential to influence and trigger shifts in users’ mobility behaviors. In this subsection, we summarize events into two categories: global and local events (see

Table 3. Event list.

Event	Date
Global	—
Russian troops were sent to Ukraine.	21 February 2022
1st-round sanctions against Russia.	23 February 2022
2nd-round sanctions against Russia.	25–28 February 2022
3rd-round sanctions against Russia.	2–9 March 2022
EU adopted temporary protection scheme for persons fleeing the war in Ukraine.	4 March 2022
4th-round sanctions against Russia.	15 March 2022
5th-round sanctions against Russia.	8 April 2022
Natural gas prices surged due to the suspension by Russia.	26 April 2022
Global	—
The war reached the outskirts of Kyiv.	27 April 2022
Russia imposed sanctions on the European subsidiaries.	11 May 2022
6th-round sanctions against Russia.	3 June 2022
More than 5.2 million refugees from Ukraine had been recorded across Europe.	4–7 July 2022
7th-round sanctions against Russia.	21 July 2022
EU suspends visa facilitation agreement with Russia.	9–12 September 2022
Both Nord Stream 1 and 2 gas pipelines ruptured.	26 September 2022
8th-round sanctions against Russia.	6 October 2022
The Council set an oil price cap for oil.	3 December 2022
9th-round sanctions against Russia.	16 December 2022
Warsaw, Poland	—
Poland opened its door to refugees from Ukraine.	30 March 2022
Poland provided resources to refugees.	1 April 2022
Polish agriculture minister claimed that Ukraine could route grain exports through Poland.	15 May 2022
Warsaw tourist population dropped as the refugee population increased.	28 June 2022
Missile struck in Poland.	15 November 2022
Discovered a large Russian missile inside Poland.	16 December 2022
Paris, France	—
National solidarity effort for Ukraine: donations of equipment and emergency vehicles.	22 March 2022
1st-round delivery of emergency medical aid.	21 April 2022
2nd-round delivery of emergency medical aid.	28 June 2022
France played its full role in hosting refugees, including financial and educational support.	28 October 2022
Berlin, Germany	—
Tens of thousands protest in Berlin against the war.	26 March 2022
Over 369,000 refugees from Ukraine registered in Germany.	22 April 2022
Support Ukraine with resources.	26 April 2022
Supply Ukraine with light weapons.	10 May 2022
Supply Ukraine with heavy weapons.	20 June 2022

). Global events refer to significant occurrences or incidents related to the Russia–Ukraine war. Local events refer to incidents that are specific to a particular locality or region (i.e., Poland, France, Britain, and Germany).

3. Performance Metrics

3.1. Gyration and Travel Distance

The S-T check-ins of a user may exhibit sparsity, but they can also show dense concentration within specific regions. These check-ins provide valuable travel-related information, including trip duration, the frequency of user visits to specific locations, and the user’s gyration, which approximates their travel distance. Mathematically,

$$\begin{array}{c}\hfill g_i=\sqrt{\frac{{\sum }_{l=1}^k{(r_i^l-\overline{r_i})}^2}{k}},\overline{r_i}=\frac{{\sum }_{l=1}^k{r}_i^l}{k},i=1,\dots ,N\end{array}$$

(1)

where $g_i$ is the gyration of user i. The coordinates $r_i^l$ represent the l-th coordinate in the user’s spatial–temporal (S-T) records. Each coordinate $r_i^l$ consists of the geo-location information collected from social media, specifically the latitude and longitude values. The index l ranges from 1 to k, representing the number of S-T records for user i. The centroid of user i’s coordinates, denoted as $\overline{r_i}$, is calculated as the average of all the user’s coordinates.

To obtain the daily travel distance for each individual user, we use the Haversine formula [22] to calculate the distance between two consecutive check-in records based on latitude and longitude coordinates. This distance can be computed using the “Geodesic” module from the Python library “Geopy”.

3.2. Change-Point Detection

To determine whether changes in users’ gyration align with global or local events, we employ change-point detection techniques to identify abrupt shifts or transitions in the time-series data. We implement the change-point detection method (CDM) from Facebook Prophet (FBP), a robust open-source forecasting tool for time-series data that effectively handles missing data and shifts in trends [23,24]. FBP excels in analyzing time-series data over various intervals—daily, weekly, monthly, and yearly—by detecting trends, abrupt changes in patterns (regime shifts), seasonality, and the impact of holidays. The core of the FBP model is a decomposable time-series model, which integrates three primary components: trend, seasonality, and holidays. The model is formulated as follows:

$$y\left(t\right)=g\left(t\right)+s\left(t\right)+h\left(t\right)+{ϵ}_t.$$

In this equation, $g\left(t\right)$ represents the trend function, capturing the non-periodic variations in the time series. The $s\left(t\right)$ component represents the seasonal changes, while $h\left(t\right)$ accounts for irregular holiday effects. The term ${ϵ}_t$ is an error component, representing random fluctuations expected to follow a normal distribution.

By using this method, we are able to produce a forecast line situated at the median of the 80% confidence interval. Additionally, we identify significant breakpoints on this forecast line with notations “Estimated Trend” in the plots in Section 4 using the CDM of FBP [25]. This approach ameliorated the sparsity and instability inherent to our initial social media dataset.

The Prophet algorithm identifies change points by initially specifying an extensive set of potential change points where the rate may alter. In our application, we prioritize the top 10 records based on the magnitude of their rate changes, specifically focusing on those with the largest absolute values. Subsequent results highlight the change points detected throughout the original data timeline, supplemented with a forecast spanning an additional month.

3.3. Granger Causality Test

The Granger causality test is a statistical method to determine whether one time series can predict another [26,27]. This test is based on the concept that if variable X influences variable Y, then the historical data of X should enhance the accuracy of predicting Y, more so than relying on Y’s own historical data. Important elements of this test are its suitability for time-series data recorded at regular intervals, such as daily or monthly, the importance of selecting a correct amount of past data points (lags) for the analysis, and the need for both data series to be stationary, indicating that their statistical characteristics do not vary over time. To ensure stationarity in our time series, we conducted both the augmented Dickey–Fuller (ADF) Test [28] and the Kwiatkowski–Phillips–Schmidt-Shin (KPSS) Test [29]. If any data series was found to be not stable, we used the difference method to adjust it and make it stationary. This test is aimed at detecting predictive relationships, rather than direct cause and effect, and usually involves using a vector autoregression (VAR) model [30] to assess the importance of previous data points in the series.

4. Case Study

This section presents a travel pattern analysis of three cases, Warsaw, Paris, and Berlin, with the aim of understanding how users’ gyration and travel distance in Europe changed after the Russia–Ukraine war.

4.1. Warsaw

We first look into the travel pattern in Warsaw. Figure 3 displays the number of check-ins, number of users, average number of check-ins, average daily gyration, and average daily distance throughout 2022. The red vertical line marks the onset of the war. The yellow, purple, gray and orange lines indicate different categories of events, including oil and gas prices, safety, refugee policies and sanctions. The change points are indicated by the red and green points. We will now delve into details of these plots.

Figure 3c depicts the temporal evolution of daily gyration and travel distance. The x-axis represents the timeline, while the y-axis corresponds to the average gyration/distance measured in miles. A notable change point was observed in the gyration curve around 29 March 2022, initiating a declining trend. The distance curve exhibited a similar declining trajectory. This can be explained by the acceptance of refugees on 30 March, which likely raised public safety concerns. The average number of check-ins (Figure 3b) also showed a distinct change point, followed by subsequent data indicating an increasing trend in social media activity. On 26 April 2022, we identified change points in the daily gyration and travel distance curves, along with a decreasing trend in the number of check-ins. This corresponds to the surge in gas prices, indicating that the price increase leads to higher travel costs and a reduction in travel frequency among individual users. We identified another change point on 1 December 2022. This observation aligned with a concurrent change point detected in the user metric curve. The daily distance metric displayed a significant increase, while the number of users and the daily gyration witnessed a slight decline. This can be attributed to the decision made on 3 December 2022, when the Council instituted a price cap on oil, benefiting travel mobility and leading to an increase in daily travel distance.

4.2. Paris

We now look into the travel pattern in Paris. Figure 4a depicts the trends in the number of check-ins and users. Compared to the trend observed in Warsaw, the number of check-ins and users in Paris does not exhibit significant changes following the onset of the war on 21 February 2022. Figure 4c demonstrates the curve of daily gyration and daily distance in Paris. On 2 May 2022, both daily gyration and distance metrics for Paris indicated a mutual decline. This trend correlates with a surge in natural gas prices due to Russia’s supply suspension on 26 April. When travel costs went up, the number of users increased, suggesting a broader awareness or usage of social media, yet the overall check-ins remained stable. This led to a decrease in the average number of check-ins per user, an understandable adjustment as individuals became more selective in their travels due to increased costs. Around 5 July 2022, a pattern emerged where gyration increased while the distance traveled decreased. This suggests that while travel frequency was reduced, the scope or area of movement broadened. On 27 September 2022, a notable divergence was observed between the metrics of distance and gyration in Paris. Specifically, distance exhibited an increase, while gyration correspondingly declined.

This deviation can be directly linked to the disruption caused by the rupture of both the Nord Stream 1 and 2 gas pipelines on 26 September. The subsequent disruption in fuel supply ostensibly led to a surge in fuel prices. Consequently, Parisians might have opted for extended yet infrequent commutes to optimize their fuel consumption. This strategy involves amalgamating multiple tasks into singular journeys, thus lengthening each trip but diminishing the aggregate number of excursions, as indicated by the reduced gyration.

Furthermore, the ascendant trajectory in the metrics for user count, check-ins, and average check-ins underscores an enhanced public engagement and heightened awareness towards events that significantly impact their daily lives. Note that there exist instances wherein the change points do not align with the event chronology. For instance, a change point was discerned around 12 April 2022 across three metrics: number of check-ins, users, and average check-ins. The solitary war-related event proximate to this date was the imposition of the fifth round of sanctions against Russia on 8 April. Notwithstanding this, all three metrics manifested a declining trend. This observation intimates the existence of exogenous variables influencing users’ mobility behaviors, transcending the scope of war-related events contemplated in this study. Delving into the causal nexus governing travel patterns and pinpointing further contributory factors constitute promising avenues for subsequent research. Another conceivable rationale is the populace’s relative indifference towards events devoid of immediate personal implications.

4.3. Berlin

Figure 5 illustrates the check-in, gyration, and distance performance in Berlin. We focus on two events: the energy supply cut by Russia and the Nord Stream pipeline sabotage [31]. Regarding the gyration of users, a change point is observed in early May, aligning with the event of Russia’s suspension of the energy supply and the subsequent surge in gas prices on 26 April. It is shown that users’ daily gyration, distance, and average check-ins in Berlin experienced a decrease during this period, corresponding to the surge in gas prices. However, the other event, the Nord Stream pipeline sabotage on 26 September, did not exhibit any clear change points in the trend of users’ check-ins, daily gyration, or distance. Overall, the surge in gas prices resulting from the energy supply cut by Russia had a substantial impact on users’ gyration and distance in Berlin. Conversely, the Nord Stream pipeline sabotage did not result in any observable changes in users’ travel patterns in the city.

We summarize our findings about users’ travel patterns and the Russia–Ukraine War:

• War-related events may contribute secondary impacts to users’ travel patterns. One of the secondary impacts observed in this study was the influence of the war on users’ gyration through factors such as the surged price of gas. As a result of the war, there were disruptions in the gas supplies, which led to an increase in gas prices. This increase in prices subsequently affected users’ travel patterns and behaviors. This indicates that economic factors from the war, such as changes in fuel costs, can negatively influence users’ travel frequency and distance. We also performed Granger causality testing on two types of crude oil prices, including West Texas Intermediate (WTI) crude and Brent crude, versus gyration. The results are shown in

  Table 4. Granger causality testing.

  Gyration	WTIx	Brentx
  Warsaw gyrationy	0.3790	0.3441
  Paris gyrationy	0.0155	0.0138
  Berlin gyrationy	0.1830	0.1853

  . The values in the table are all p-values. The observed p-values for the pairs $({\mathrm{WTI}}_x,\mathrm{Paris}{\mathrm{gyration}}_y)$ and $({\mathrm{Brent}}_x,\mathrm{Paris}{\mathrm{gyration}}_y)$ are found to be less than 0.05. Based on this statistical evidence, it can be concluded that WTI and Brent crude oil prices exhibit Granger causality in relation to Paris gyration with a 5% significance level.
• The impact of global and local events regarding the Russia–Ukraine war on users’ travel patterns can differ based on their own scope. Global events, such as surged gas prices, can directly affect transportation systems, leading to more widespread changes in gyration patterns in major cities. On the other hand, local events, like the missile strike in Poland, have a more limited impact on users’ gyration patterns, mainly affecting the specific area where the event occurred. Understanding the impact of global and local events on users’ gyration is crucial for comprehensively analyzing and interpreting changes in mobility behaviors.

5. Conclusions

What is the impact of the Russia–Ukraine war on human mobility patterns? Rather than using individual private trace data, here, we use unintrusive, low-cost social media check-in location data to infer such an impact. In this work, we focus on understanding the impact of the Russia–Ukraine war on human mobility patterns, particularly in European cities. We utilize Twitter check-in location data to analyze the trends in check-ins and gyration across different cities. Our findings reveal that the war and related events had varying effects on users’ mobility behaviors in different regions. In Warsaw, the number of check-ins significantly increased after the war. Additionally, the gyration in Warsaw decreased by 50% after the gas prices increased by 54%. In Paris and Berlin, the impact of the war on check-ins and gyration was less pronounced, with other factors such as gas prices playing a secondary role.

In our future research, we aim to look into additional factors that can influence these patterns. One aspect to explore is the impact of refugee policies on users’ mobility behaviors. We will study individual traces obtained from check-in data to examine how policies implemented by different countries and regions, such as the European Union’s protection plans or specific refugee acceptance programs, shape the mobility patterns of individuals affected by the war. We will also delve deeper into the analysis of the causal relationships among the war, refugee policies, gas price, and human mobility.