Who, Where, Why and When? Using Smart Card and Social Media Data to Understand Urban Mobility

Abstract

This study describes the integration and analysis of travel smart card data (SCD) with points of interest (POIs) from social media for a case study in Shenzhen, China. SCD ticket price with tap-in and tap-out times was used to identify different groups of travellers. The study examines the temporal variations in mobility, identifies different groups of users and characterises their trip purpose and identifies sub-groups of users with different travel patterns. Different groups were identified based on their travel times and trip costs. The trip purpose associated with different groups was evaluated by constructing zones around metro station locations and identifying the POIs in each zone. Each POI was allocated to one of six land use types, and each zone was allocated a set of land use weights based on the number of POI check-ins for the POIs in that zone. Trip purpose was then inferred from trip time linked to the land use at the origin and destination zones using a novel “land use change rate” measure. A cluster analysis was used to identify sub-groups of users based on individual temporal travel patterns, which were used to generate a novel “boarding time profile”. The results show how different groups of users can be identified and the differences in trip times and trip purpose quantified between and within groups. Limitations of the study are discussed and a number of areas for further work identified, including linking to socioeconomic data and a deeper consideration of the timestamps of POI check-ins to support the inference of dynamic and multiple land uses at one location. The methods and metrics developed by this research use social media POI data to semantically contextualise information derived from the SCD and to overcome the drawbacks and limitations of traditional travel survey data. They are novel and generalizable to other studies. They quantify spatiotemporal mobility patterns for different groups of travellers and infer how their purposes of their journeys change through the day. In so doing, they support a more nuanced and detailed view of who, where, when and why people use city spaces.

1. Introduction

Understanding urban flows and dynamics is important for uncovering hidden knowledge in spatial and social systems. For example, Batty [1] argues that cities are built around flows of money, information, resources, etc. as well as people across urban spaces. Exploring how individual citizens move around urban spaces can potentially shed new light on both urban space characteristics and, critically, their dynamics and complexities [1,2].

Knowing how, where, when and why people travel in cities, particularly on a large and comprehensive scale, remains a challenge for researchers. Traditional travel surveys [3,4,5,6] are simply not responsive enough to capture the dynamics of population flows within cities and critically how patterns of movement change temporally as well as spatially. Transport system smart card data (SCD) are passively collected by automated fare collection systems in stations or on vehicles. They record individual-level details of where and when travellers enter (tap-in) and leave (tap-out) the transit system. They capture the dynamics of individual mobility within the city and provide opportunities to generate new insights into travel flows and mobility behaviours. However, such data contain no information on traveller socioeconomic status or trip purpose [7]. New forms of micro-level (big) data, such as from social media, have been found to contain rich information about place semantics and individual interactions with the physical world [8]. Combining such information with SCD presents an opportunity to generate a more holistic picture of urban flows through inference of where, when and why individuals move through cities. These understandings can also benefit the related urban and infrastructure planning, for example, contributing to the development of “liveable city” [9].

In this context, the aims of this paper were (i) to link metro SCD with land use inferred from social media check-ins at points of interest (POIs), thereby (ii) to generate travel profiles from their origin and destination and from the time and day of travel, and to infer trip purpose, and finally (iii) to analyse travel flows of different groups and sub-groups of travellers to generate new insights into how individuals interact with and use urban space. The paper is organised as follows: Section 2 presents an overview of the issues around understanding urban mobility, with Section 3 presenting the data. The methods are presented in Section 4, analysis and results are in Section 5, with a critical discussion on limitations and areas for further work given in Section 6 and Section 7.

2. Behaviour from Smart Card Data

The analysis of mobility patterns within public transit systems can reveal new insights into the spatiotemporal features of daily urban life. An improved understanding of the mobility patterns of transit riders from different socio-economic backgrounds can support the evaluation of different aspects of current public transit services by authorities and policy makers. This allows, for example, targeted marketing strategies, decision making to improve services and explorations of the resilience and efficiency of transport infrastructures.

Historically, such activities have been informed by travel behaviours research based on questionnaires and travel surveys [3,4,5,6]. Whilst survey data commonly contain personal demographic and socioeconomic details of survey subjects, they have a number of drawbacks. First, the representativeness and generalizability of the information from surveys may be limited, with a small number of respondents typically sampled. They may not be conducted at the same places and can have short temporal currency, particularly in cities that have been subject to rapid urbanisation over recent decades [10]. Travel surveys may fail to adequately represent these situations. For these reasons, there has been an upsurge in research interest exploring the opportunities afforded by the many new forms of big data, including social media travel card data and social media.

Smart card data (SCD) are event-triggered. Transactions are recorded only when the traveller swipes their card to board a vehicle or access a station. SCD have been used by researchers to investigate patterns of urban flows, including commuting, mobility and travel areas [11,12,13,14,15]. These studies have focussed on identifying the spatiotemporal patterns within the SCD in order to inform and support transportation planning. Such studies are plentiful, and typically they evaluate the spatiotemporal patterns of trips through the transit system [14] to quantify and predict individual mobility [14] to examine route choices [14], the scales of regular and explicable travel behaviours [16] and temporal changes in the spatial structure of urban movement [12]. Comprehensive reviews of the technologies, applications and methodologies of SCD analyses and the evolution of thinking in this area are provided by Bagchi and White [17], Pelletier et al. [7] and by Li et al. [18].

One of the main difficulties experienced in research and analyses of SCD is how to link the observed variations in urban flows and spatiotemporal dynamics with individual socioeconomic attributes and thereby infer the purpose of trips. Some studies have been able to classify travellers into different groups and have analysed these separately in order to gain a better understanding of cardholders’ travel behaviour. For example, Huang [19] studied the diversity of spatial and temporal mobility patterns of different age groups (child/student, adult and senior citizen). Wang et al. [20] and Long et al. [6] analysed university students and those making unusually long, early, late or daily trips (“extreme transit commuters”). Other research has identified peak travel times for specific groups, such as students [19]. Although these studies included demographic dimensions and have advanced understanding, they all concluded that a lack of socioeconomic and demographic details, and in particular an absence of data on the purpose of journeys, presented a major barrier to more in-depth and useful studies. Others have sought to infer such characteristics from the time, origin and destination of trips, but inferring trip purpose presents a challenge. In many cases some kind of service area or catchment has been used, defined as either a buffer (fixed distance or isochrone) around metro stations or administrative polygons [21]. Such areas have also been used to characterise the origin or destination areas, frequently through land use designations. For example, Wolf [22] suggested that matching land use information with trip origin and destination could give greater insight into individual motivations, providing context for specific trips and thereby potentially supporting inferences of trip. Lee and Hickman [23] and Devillaine et al. [24] used a combination of decision tree and heuristic rules to infer trip purpose from trip temporal characteristics, socioeconomic and land use information. Their method is highly dependent on the duration of activity, which is based on the assumption that users do not use any other transit modes in their trips. This assumption is not likely to be true for all travellers, especially occasional travellers. The work of Medina [25] combined household travel surveys and high quality public transport data for inferring bus and metro trip purpose of going home, to work or study. Despite its effectiveness, it may have a shortcoming in applicability for other places. Not many cities in developing countries (e.g., China) conducted household travel surveys regularly, and their buses do not often contain both boarding and alighting information to construct a bus and metro travel chain, as proposed in [25]. Liu et al. [26] studied the dynamics of the inhabitants’ daily mobility patterns using smart card data in Shenzhen. They identified morning metro tap-in as being close to specific residential areas and afternoon tap-in close to large working areas using detailed examples. However, this research only analysed specific locations, provided no system-wide analysis and ignored other potential land use types.

Another study in Shenzhen [27] sought to link bus, metro and taxi trips using a spectral clustering approach to analyse transit mode. This was used to delineate five urban space categories, for which the urban function was manually inferred, and to suggest mass transit patterns from the category socioeconomic characteristics.

A number of similar or improved methods (e.g., probabilistic model [28]) have been proposed [29,30], combining detailed GPS tracking data with land use information to detect both transportation mode and trip purposes. Whilst providing a richer overview of movements, this approach is limited by the number of individual study participants. Nonetheless, land use describes socioeconomic activities and provides a prism by which to infer trip purpose.

The problem encountered by previous research is that land use at any given origin or destination is unlikely to be unique—multiple land uses co-exist in space and time (see Fisher et al. [31] for a full treatment of this issue). Thus, although a number of methods have been developed for inferring trip purpose from land use [22,23,29,30], they all face the same problem of how to identify the important land use entities in different parts of the transportation system. Analysis of social media check-in data can provide an indication of this.

Social media data analysis can be used to provide an understanding of local sentiment, as well as where individuals go and why [8]. Much land use-related information is also recorded both directly and indirectly in social media. Indirect information may be through the description of activities that are being undertaken, and direct land use information is available through point of interest (POI) check-ins. These record the presence of social media users at specific labelled locations. POI data have been found to enrich spatiotemporal semantic information in analyses of urban space [8,32,33] by supporting inference of people’s activity in physical space. POI or parcel level land use data have been used to enrich information around origins or destinations [22,23,24,30]. In previous research, POIs or land use have been assumed to have the same potential to originate or attract trips, a simplification which leads to a bias in representing trip purpose inference [22]. For example, a large residential POI may be more important than a small shopping mall in originating trips, and this should be reflected in different weights within a trip purpose inference model.

In summary, socioeconomic information can support deeper understandings about trip purpose, thereby providing richer analyses of urban flows and dynamics. Some research has shown it is possible that trip purpose can be inferred from trip pattern and regularity [34]. Land uses at trip origins and destinations allow a degree of socioeconomic and purpose characterisation. Where lacking or where the land use is uncertain, it can be inferred from POI check-ins in social media data. Current studies using SCD normally focus on general mobility behaviours, such as travel frequency, travel distance and regular origin-destination (OD) pairs. The potential for contextual information derived from low-cost social media has not been fully exploited. Similarly, much of the literature focuses on methods to infer an individual trip’s purpose and fails to shed light on the overall trip purpose pattern in the whole transit system. The research presented in this paper addresses these and a number of other gaps: Social media POI check-in data are used to quantify POI weights, allowing a more accurate description of land use information to be derived, and changes in trip purpose patterns for individuals are evaluated to shed light on when and why different people travel within the city.

3. Study Area and Materials

3.1. Study Area

Shenzhen is a city region with a resident population of around 12 million. It lies just to the north of Hong Kong, covers an area of around 2000km² and is part of a broader region with significant employment in financial services and high-tech industries. Rapid urbanisation and urban transport development resulted in a metro system with five lines and 118 stations by 2014. The metro lines connect with the Hong Kong metro system at Futian Checkpoint Station and Luohu Station. Figure 1a,b shows a map of the Shenzhen metro system with community administrative zones and the population density of each zone. The community boundaries were provided by the Future Transport Lab (Shenzhen). The main urban area is in the mid-southwest, with higher population densities, and the northern and eastern part of Shenzhen are suburban with low densities.

3.2. Data

3.2.1. Shenzhen Public Transport

Metro smart card trip data for Shenzhen were obtained for the period 9 June 2014 to 13 June 2014 (Monday to Friday) from the Transport Commission of Shenzhen Municipality and the Asia and Pacific Mathematical Contest in Modelling Committee. A total of around 8 million metro trips were analysed in this study, recording the movement of around 2 million individually registered smart card users. The trip data included attributes describing the user ID, trip time (tap-in and tap-out timestamps), price (full or discounted according to different card type), the tap-in and tap-out station names, the metro train ID and the metro line name. The category of travellers (e.g., student, elderly or disabled) can be identified through the different discounted ratio of travel price. Morning commuters were distinguished from other adult travellers if they repeatedly travelled in the morning (between 6:00–11:00 a.m.) for at least four days in the five weekdays in this study. The number of commuters making metro trips in the study period was 443,650, and they made a total of 3.45 million metro trips in the study period.

Other data detailed the location (latitude and longitude) of the metro stations, which enabled origin and destination locations to be determined.

3.2.2. Social Media Data

Social media captures an unprecedented level of detail on human activity. The biggest micro-blogging platform in China is Weibo. It encourages users to check in to local POIs when they are at those locations. Weibo provides aggregated check-in data, which contain attributes describing the POI ID, its name, address, latitude, longitude, category name and the total number of check-ins. The aggregated data record only the total number of check-ins rather than individual check-ins. It should be noted that the data do not include any check-in timestamp information. The POI category allows the POI-related activities to be inferred and classified. For example, theatres can be classified into “leisure” activities. The total number of check-ins at each POI can be used to provide a weight to the activities therein and their relative importance, for example in analyses seeking to model local spatial characteristics.

The Weibo check-in data used in this research were made available to this research, covering the period June 2011 to November 2014. The number of POIs in Shenzhen reached around 70,000, labelled with 221 categories, and the total number of check-ins was around 1.5 million. The 221 categories were reclassified into six land use types, as shown in

Table 1. Reclassified point of interest (POI) categories.

Reclassified Categories	Examples
Commercial	Retail, restaurant, life service (e.g., car washing, barber), etc.
Leisure	Sports, entertainment (e.g., karaoke), scenery spot, etc.
Residential	Residential community, residential building, etc.
Working	Office and industrial places.
Education	School, library, research institution, etc.
Other	Landmark, communal facilities (e.g., public lavatory), etc.

. Some POIs could not be classified, for example landmarks, and these were reclassified into the class of “other” and were not used in the analysis. A point-in-polygon operation was used to determine the proportion of each POI type in each zone (Figure 2). Zones mainly contributing to commercial and leisure activities are relatively balanced and all over the city, but education zones are mostly in south and northeast. The working zones are also centred at the middle and north of the city.

Not all the data that are the basis for the results in this paper are publicly available. Some data and code that produce the results are available from University of Leeds data repository: https://doi.org/10.5518/599.

4. Method

4.1. Temporal Trip Purpose Inference

The travel flows between metro stations from tap-ins and tap-outs provide limited information about how people interact with urban spaces—it is difficult to infer spatiotemporal patterns in trip purpose directly from travel records. This is an inherent shortcoming of anonymised SCD generated from automated fare collection systems, which lack information about individual socioeconomic activity that may be used to support inference about trip purpose. To overcome this, aggregated Weibo check-in data were used to infer and weight potential land uses and activities in the areas around each metro station. For each metro station, a service catchment area was defined by a 2500m buffer. This distance was used to represent a walking time of less than 30 minutes and a cycling time of around 10 minutes. Please also note that the check-ins do not contain individual timestamp information—only the aggregated number is used.

For each catchment area, the number and type of different POI check-ins (residential, working, commercial, leisure, education) were determined. Figure 3a shows the proportion of different type of check-ins using boxplot, and the correlations between different type of POI check-ins number and metro trip amount are illustrated in Figure 3b. Overall, residential POIs contributed to most of the check-ins, followed by retail and working. Education has the lowest proportion on average (Figure 3a). All types of check-ins are found to have positive correlations with metro trip amount (Figure 3b), working presents the highest correlation, followed by commercial and residential.

The trip purpose was then inferred by comparing the proportion of different kinds of POI check-in at traveller origin and destination catchment areas. For example, a trip from an area with predominantly residential POI check-ins to a one dominated by leisure POIs check-ins (e.g., an area with lots of check-ins at parks) is indicative of a passenger travelling from home to take part in leisure activities. Other trip purposes can be inferred from the change in POI check-in types between OD (Origin and Destination) metro station service catchments in a similar way.

People use the transit system to reach different places, for different purposes at different times of the day. In order to examine temporal variations in trip purpose, the operational metro service hours (6:00–23:00) were sliced into hourly intervals, and the aggregated total numbers of different categories of POI check-ins for each trip’s origin and destination metro station service catchment could be accumulated and compared.

To examine the differences between the land use groups further, a change rate was defined (Equation (1)) to support a clear and deeper analysis of the temporal changes in trip origin and destination,

$${\mathrm{R}}_i=\frac{O_i-D_i}{\max \left(O_i,D_i\right)}\times 100\%$$

(1)

where ${\mathrm{R}}_i$ represents the change rate of land use category I, $O_i$ is the ratio of land use category i at the origin, and $D_i$ is the ratio of land use category i at the destination. The change rates provide an aggregate measure of how people travel from between land use areas in each time period. For example, if the rate for residential is positive, and the rate for working is negative, then this suggests that travellers are leaving work and going home.

4.2. Traveler Division Based on Boarding Profile

Understanding the travel flows of different types (e.g., students, elderly) of travellers helps uncover new insights into how individuals interact with and use urban space. However, sub-groups of behaviours exist within the broad-scale groups that exhibit more nuanced spatio-temporal characteristics. The sub-groups are worthy of exploring because different travel behaviours are considered to be associated with travellers’ socioeconomic background [6] (e.g., income and education level). Dividing transit users into detailed categories and understanding their travel patterns may help to improve transport services. Rule-based approaches have been used within the literature [6,23,24] to extract certain travel behaviours (e.g., travel in early morning), but this relies heavily on formulated and arbitrary rules. To overcome it, unsupervised learning can be applied to categorise travellers based on their travel behaviour and help to uncover the sub-groups and their behaviour in transit systems. To do this, boarding time profiles were created for each card holder, in which the hourly trips were counted. For example, Figure 4 shows a hypothetical boarding time profile. This passenger made three trips between 6:00–7:00 on Thursday with only two trips made in the other days. The profile is, in essence, a matrix of 5×24, with each number representing the count of tap-ins in 1-hour intervals.

To evaluate the similarity between individual boarding time profiles, the average (mean) and variance value of boarding numbers for each time slot, were calculated for each user. This generates a total of 48 daily measures, which were used to characterise the temporal travel pattern of each individual i and to calculate a boarding time profile, $C_i$, denoted as

$$C_i=\left[A_{i,1},A_{i,2},\cdots ,A_{i,24},V_{i,1},V_{i,2},\cdots ,V_{i,24}\right]$$

(2)

where $A_{i1}$ is the mean at the first-time slice and $V_{i1}$ is the variance at the first-time slice. After extracting the $C_i$, unsupervised learning (e.g., k-mean clustering) can be applied based the feature collection, thus dividing travellers into sub-groups.

5. Analysis and Results

The analysis sought to explore the temporality of trip patterns for different groups of users, comparing students against all travellers (Section 5.1), to infer trip purpose from the land uses associated with trip origins and destinations, comparing commuters, students and all travellers (Section 5.2) and to create profiles of the spatio-temporal behaviours of different sub-groups (Section 5.2.3).

Figure 5 shows the flow map of trips in Shenzhen metro system, indicating the origin (tap-in) and destinations (tap-out) for all travellers. The popular stations are mostly located in south central of Shenzhen. Chegongmiao station is the most import hub in the network, with the largest number of tap-ins and tap-outs among all stations.

5.1. Temporal Mobility Analysis

Examining trip frequency and trip temporal density of different groups of users provides insight into their varying mobility patterns. They are also associated with underlying different trip purpose. In order to understand temporal mobility patterns, each day was divided into 48 time intervals, and the number of metro trips in each interval were counted. Figure 6 shows the trip tap-in counts for all travellers and for school students during weekdays. It reveals a degree of regularity of trip patterns. For all passengers, two peak hours on weekdays are evident, one from 7:30 to 8:30 and another from 17:00 to 18:30. First, the pattern for student trips diverges from these general trends in a number of ways. The morning peak is one hour earlier (from 6:30 to 7:30), this is because of earlier school start times (commencing at 7:30–8:00). Second, the number of all travellers’ trips on the Friday evening is slightly higher than on other days. Two small student travel peaks emerge at lunch time, perhaps due to them returning home to eat, with the first peak travelling to home and the second going back to school. The lunch time peak for students on Friday is higher than other weekdays, which is similar to that found in studies of Singapore metro travellers [11]. Additionally, students make more trips in the afternoon and in the evening on Fridays, with the peak volume nearly twice that of other weekday peaks and commencing approximately one hour earlier (16:00) than on other days (17:00). The potential reason for this abnormal bump is examined in Section 5.2.3.

5.2. Trip Purpose Pattern Analysis

5.2.1. Temporal Trip Purpose of Adult Travellers

Supplemental contextual information for travel flows can contribute to an understanding of trip purpose [8,32,33]. The analysis of temporal trip purpose utilised the methods described above. Recall that commuters were defined as those transport system users who travelled in the morning (6:00–11:00) for at least four out of the five weekdays. Their trips were compared with trips by all travellers, as shown in Figure 7. This indicates that the main changes in origin and destination proportions are driven by working and residential land uses during normal rush hour periods for both traveller groups, with similar relative increases in education, and that the volumes of trips from and to commercial land use remain constant through the day. However, it is difficult to infer important differences between these groups. To better interpret the temporal variation in travel purpose, the change rate ${\mathrm{R}}_i$ is calculated for the two broad-scale groups at each time interval.

Figure 8 shows the change rate for different kinds of land use inferred from POI check-ins in origins and destinations. Figure 8a allows the overall trip purposes for travellers to be inferred. People leave residential areas and to go to working and education areas in the morning rush hour, between 7:00–12:00, with the greatest value at 7:00. The trends decrease and converge at 11:00. Between 12:00 and 14:00, the overall changes of trips between different land use areas are near to 0, with the rate for working showing a small peak and the rate for residential a small trough, indicating that some travellers return home to eat and then go back for work. Then, from 14:00, the morning pattern is reversed, and the values for residential areas become positive while the values for working and education areas change to negative as travellers begin to leave work and school to return home. There are some increases for commercial areas from 17:00 to 20:00. The change rate of leisure was negative (−12%) in the early morning, it increased and stayed steady around 0 at noon, and finally reached around 10% after 17:00. Unlike the curve of commercial, change rate of leisure stayed around 10% after 21:00, which implies that a certain number of metro travellers made their trip for leisure activities at late night, while trips with commercial purpose are approaching zero (21:00), probably due to the closing time of stores and malls (after 21:00).

For commuters (Figure 8b), the pattern is similar, but some subtle differences related to working patterns are observable. The flows from residential to working and to education are before 10:00. They have a much sharper drop from their peaks to 0 before the midday. This indicates that the shifts from residential to working are concentrated from 6:00 to 10:00 compared to similar changes for all travellers (06:00–12:00). Travellers labelled as commuters are more likely to have a regular scheduled job starting before 10:00.

Different trip purpose can be interpreted from the temporal variation in change rates of different land use types. The most significant observations are working and residential. This is due to the higher degree of work-home spatial separation than other land use types in the city. Figure 2d also indicates that working-related POI check-ins are more spatially heterogenous and mainly clustered in several zones in the south.

5.2.2. Temporal Trip Purpose of Students

Trips undertaken by students were analysed to compare trips made on Monday to Thursday with those made on Fridays. Figure 6b highlights dramatic differences in afternoon trips on Fridays. The change rate (Equation (1)) was used to unpick variations in student trip purpose on Friday afternoons and evenings. Because student trips are not related to work activities, work-related POI check-ins were excluded from the analysis. The results are shown in Figure 9 and Figure 10. These indicate that on Monday to Thursday, the change for residential dominates. Figure 9 and Figure 10 indicate that the main trip purpose is returning home from school, but during Friday’s afternoon peak hour (from 16:00 to 18:00) there is a dramatic decrease in the change rate for residential, which is negative around 16:00 (Figure 10b). In contrast, change rate for commercial becomes the highest around 16:00. The switch of position of commercial and residential indicates that the main trip purpose for students during the Friday afternoon peak hour is “consuming in commercial areas” rather than returning home as on other weekdays. The rate for leisure also increases to positive from 16:00 to 18:00, suggesting that more trips were made for the purposes of leisure activities compared to other weekdays.

5.2.3. Traveller Division and Detailed Temporal Trip Purpose

The final analysis sought to identify and compare behaviours sub-groups in terms of their patterns of travel using an example of student travellers. Student sub-groups (clusters) were identified from similar temporal travel behaviour patterns.

After creating boarding time profile and feature extraction (described in Section 4.2), the k-means algorithm was then used to cluster travellers into different sub-categories. The number of clusters was determined using the “elbow” method [35]. This determines clusters using different values of k for which a cost/evaluation measure is calculated. At some values of k, the cost/evaluation reaches a plateau for further increases in k, indicating an appropriate number of clusters. Figure 11 shows the k-means clustering using different k values. The right scree plot of Figure 11 shows the sum of squared error (SSE) when increasing k, and the left plot indicates students of different clusters in the x axis and the number of clusters in the ordinate, with the clusters represented by colours. When increasing the number of clusters, students are categorised into finer groups with more varied colours. In this study, the elbow cannot be unambiguously identified, and two present themselves: the first and most obvious one is located around k = 5 and another is located between 14 to 18. A value of k = 16 was chosen for the k-means clustering to unpick finer and more varied travel behaviour sub-groups.

The average boarding time profile for the 16 clusters is shown in Figure 12. The shading saturation (“values” in legends) indicates the probability of boarding for each sub-group for each time period and can also be interpreted as how many trips are made by travellers on average. Cluster 9 and 11 have similar travel patterns, but cluster 9 students travelled earlier in the morning and boarded later in the afternoon than cluster 11. This reflects differences between primary, middle and other schools, with middle schools having earlier and longer school days. Thus, it is reasonable to assume that cluster 9 represents middle school students while cluster 11 is primary school students. Clusters 3, 5, 12, 14 and 16 represent 15.77% of the students, and all made trips during lunch time. The most interesting group is cluster 7 (36% of all students), who travelled mostly on Friday afternoon at peak hours, making very few other trips at other times. The trip purpose of students in cluster 7 was examined in more detail (Figure 13), however, this is complex and mixed, with a slightly stronger trip purpose to residential (potentially going home).

The ratio associated with commercial for cluster 7 students rises significantly from 26.99% to 40.60% during 16:00–17:00 and is then steady during 17:00–19:00 compared to a drop from 40.75% to 35.21% during 17:00–18:00 and a drop throughout the whole period for other students (Figure 13b). In addition, the ratio associated with leisure for cluster 7 students increases during 16:00–18:00 then shows a decrease after 18:00 compared to a steady decrease for other students. The results indicate that many of those in cluster 7 take trips for commercial and leisure-related activities, which is different from other students who are mainly going to residential places.

6. Discussion

The results of this analysis infer travel behaviours for different groups of metro system users based on the temporal and spatial patterns of their trips, as recorded in smart card data (SCD) and linked to social media data. Land use at origin and destination locations were used to infer trip purpose, providing details and explanations of trips for different users at different times. The SCD allowed different groups of users to be identified based on their fare reductions (students), travel times (commuters) along with the land use derived from social media point of interest (POI) check-ins. These groups were further explored to identify different clusters of student travellers based on the temporal profile of their metro use.

This analysis of new sources of big data to examine travel behaviour addresses the obvious drawbacks and limitations of traditional travel survey data. The work generates similarity metrics for individual traveller profiles based on average trip times and their variance and is generalizable to other studies for user classification based on usage/interaction pattern. These may benefit from applying the per land use change rate approach and from transforming the travel data into “traveller profiles” to determine clusters of users. These could also be applied on repetitive timescales, for example. Moreover, this work used POI data to infer land use-related contextual information around metro stations. Potential land use activities were weighted by quantifying the number of POI check-ins at each POI to eliminate potential bias of treating all POIs equally. The proposed “change rate” measure is capable of supporting related visualisation analysis by providing clear trip purpose interpretation from flow-associated POIs.

Both the smart card data and social media check-in data may be subject to sampling bias, with the impacts of biases in the POI data potentially more serious. Here, only 70,000 POIs were used, a very limited proportion of the total number of POIs in Shenzhen, and their time stamps were between 2011 to 2014 and potentially subject to land use changes in that time. Therefore, a post-hoc validation exercise was undertaken to quantify the potential for bias in the Weibo check-in data. A sample of 1000 Weibo POIs were overlaid with reclassified Baidu map data. Here, the Baidu API was used to extract Baidu POI information in June 2018. The overall correspondence was 0.87, and the Type I errors rates for commercial, education, leisure, working and residential were 0.19, 0.14, 0.17, 0.04 and 0.17, respectively. These indicate the proportion of times a POI land use label used in this study was incorrect (false positives). These error rates suggest that the broad findings about trip destination and purpose are reliable but with varying degrees of uncertainty. The rapid development of Shenzhen may contribute to a high-speed change in urban land use of different areas, resulting in more uncertainty of using aggregated POI check-in data over long period. Ideally, using temporal matched contextual information can lead to a more accurate interpretation of local land use and trip purpose inference.

There are a number of limitations to this study and areas for further work. First, the analysis and identification of traveller segments, temporal variation in mobility patterns and trip purpose evaluated commuters and students against all travellers. These groups could be expanded, for example by examining the socioeconomic properties of the areas from which travellers originate (in the manner of geo-demographic classifications). This would support probabilistic inferences of more nuanced and detailed trip purposes for a wider number of traveller groups and potentially would allow more precise and explanatory analyses of travel behaviours, expanding the results presented here. Such analyses would also allow the potential biases in the representativeness of SCD to be quantified, as travel smart cards may not be used equally by all social groups and offer a potential avenue to understand the sample biases associated with POI check-ins, which may over-represent particular types of activities, with, for example, people more likely to post micro-blogs while undertaking leisure activities, compared to domestic ones. Second, land use is not static, rather multiple, alternate and dynamic land use attributes may be present [31] as the socioeconomic activities associated with any given location (origins and destinations in this study) may change over the course of a day. There is a need for studies of urban flows and dynamics to accommodate the dynamic nature of the concept of land use, which may have specific functions at different times during the day. Here, the land use from the POI data was inferred by the weighted aggregate check-ins to each POI in each zone. The timestamp of individual check-ins was not considered, which would allow the land use associated with each zone to be inferred dynamically. Such temporal refinements would extend and improve analysis of trip purpose.

Smart card data in the metro system are only one kind of mobility big data that contain travel information of urban inhabitants. Other similar data include bike-sharing transactions and taxi trip record. Smart card data of the bus and metro can generally capture flows and dynamics in mid- to long-distance trips but may not be suitable for describing more localised travel and activities compared to bike-sharing data. It should be noted that some other kinds of consumer data [36] can also be used to derive mobility-related information, some examples include cell phone tracking records, social media data (e.g., geo-tagged Twitter) and retail transaction records. The enumerated data all contain their own set of shortcomings, for example, sampling bias, due to their varied attractiveness to different kinds of urban inhabitants. Cell phone tracking records are considered to have high representativeness of users and may suffer least of all from the bias problem, but they have another shortcoming—the data may fail to record every OD pair to represent travel flow because the location of the user is only tracked when the phone is being used. These data can, to some extent, reveal urban dynamics and how people interact and utilise urban space, but combining data from different sources may contribute to a more comprehensive picture of urban flows. Although varied in structure and spatiotemporal granularity, such data and derived travel flows always benefit from external contextual information. Incorporating contextual data (e.g., land use) can lead to deeper understanding of the flows and activities. The obtained insight on flows of individuals and groups of people at different time periods also reveals the complexity and diversity within city life.

7. Conclusions

New forms of smart card data present opportunities to generate new insights into the dynamics and patterns of flows within cities. This study demonstrates how combining SCD with other context-rich data, such as social media data, are able to reveal the dynamics of mobility patterns associated with urban transit systems, and these vary amongst different social groups, supporting changes in urban planning. The study has demonstrated how linking such data can be analysed to infer: (a) The behaviours of different groups of travellers, (b) the socioeconomic activities that people undertake and (c) how different groups (and sub-groups) identified in these ways are associated with different travel behaviours, trip purposes and socioeconomic activities. In so doing, this research addresses the drawbacks and limitations associated with traditional travel survey data. It develops methods that are generalizable to other studies, supporting a more nuanced and detailed view of who, where, when and why people use city spaces. The approach uses social media POI data to semantically contextualise information derived from the SCD to allow trip purpose to be inferred, to quantify spatiotemporal mobility patterns for different groups of travellers and to infer how their purposes of their journeys change through the day. Future work will extend this research to explore the links between integrated transportation trips, for example, via metro and bus, linked to use of new dockless bike schemes. Data from dockless bike sharing systems support spatiotemporal analyses of urban dynamics at a finer granularity than is possible through analysis of travel card or dock-based bike scheme. The increased uptake of station-free travel data supports enhanced understandings of urban dynamics and travel patterns over the “last mile” at much finer resolutions than has hitherto been possible.