Identifying human mobility patterns using smart card data

ABSTRACT

Human mobility is subject to collective dynamics that are the outcome of numerous individual choices. Smart card data which originated as a means of facilitating automated fare collection has emerged as an invaluable source for analysing mobility patterns. A variety of clustering and segmentation techniques has been adopted and adapted for applications ranging from market segmentation to the analysis of urban activity locations. In this paper we provide a systematic review of the state-of-the-art on clustering public transport users based on their temporal or spatial-temporal characteristics as well as studies that use the latter to characterise individual stations, lines or urban areas. Furthermore, a critical review of the literature reveals an important distinction between studies focusing on the intra-personal variability of travel patterns versus those concerned with the inter-personal variability of travel patterns. We synthesise the key analysis approaches as well as substantive findings and subsequently identify common trends and shortcomings and outline related directions for further research.

KEYWORDS:

• Travel patterns; public transport; smart card data; market segmentation; clustering; urban analytics

1. Introduction

1.1. Smart card data and the identification of mobility patterns

Human mobility is subject to collective dynamics that are the outcome of numerous individual choices. Even though individual needs and travel choices exhibit large variability and the urban and regional areas in which those they are embedded are subject to great diversity, there is evidence to suggest that human mobility manifests some recurring features throughout history and across geographies (Ahmed & Stopher, 2014). The growing availability of disaggregate mobility data enables the analysis of temporal and spatial patterns and the link between microscopic behaviour and resulting aggregate flows (Schläpfer et al., 2021). Geo-location traces are increasingly available from mobile phone and GPS data, social media posts and travel-related records such as automated fare collection (AFC) records (see Barbosa et al., 2018), giving rise to a plethora of studies analysing human mobility patterns and the decomposition thereof in the past decade. In this paper we provide a systematic review of the state-of-the-art on identifying and clustering human mobility patterns from smart card data and propose an outlook for addressing persisting challenges.

Smart card transactions offer a unique and rich source of passively collected data that enable the analysis of individual travel patterns. In the last decade, an extensive research attention has been devoted to the identification and classification of mobility patterns from smart card data. A number of review papers were devoted since the early 2010′s to smart card data analysis. Pelletier et al. (2011) reviewed the technological advancements making the analysis of smart card data possible, followed by a review of strategic, tactical and operational applications of smart card data in public transport. Another relevant work is the one by Yue et al. (2014) who reviewed trajectory-based travel behaviour studies, where smart card data is covered as one of the emerging sources enabling the analysis of human mobility patterns, replacing travel diaries and stated preferences data. Their main focus is on the properties of different trajectory data categories. In a more recent review, Hussain et al. (2021) examined recent developments in using smartcard data for the purpose of origin-destination matrix estimation. The latter – as well as related inference algorithms that are performed in order to estimate travel destinations, vehicles boarded, transfer locations and home-zones – is a pre-requisite for conducting some of the research devoted to identifying human patterns based on smart card data. In this review we seek to answer the following three questions:

• What are the key approaches and findings in relation to the analysis of individual spatial-temporal travel patterns based on smart card data transactions and in performing a market segmentation analysis?

• What are the key approaches and findings in relation to the analysis of urban and network characteristics based on individual’s spatial-temporal travel patterns as revealed by smart card data transactions?

• What are the key on-going trends, common shortcomings, knowledge gaps and related research directions?

1.2. Research strategy

Our search strategy was based on applying combinations of “Smart card data” and one or more of the following keywords when performing the bibliometric search: “Clustering”, “Travel patterns”, “Mobility patterns”, “User segmentation”, “Spatial”, “Temporal” and “Spatiotemporal”. The term “user” is used interchangeably with the terms “traveller” and “passenger” in the public transport demand analysis literature and we indeed find that these terms are used interchangeably in referring to individuals.

The databases of Scopus and Google Scholar were used to identify all relevant literature. The search, last performed on February 17, 2023, resulted in 724 papers of which 57 papers were identified as directly relevant for the scope of this analysis. Several additional studies were identified by means of a snowball search. Some of those excluded performed aggregate descriptive or statistical analysis of the spatial-temporal properties of public transport demand rather than identifying distinctive ones. Other papers excluded from this review focused on one of more of the following topics: inference algorithms, modes other than public transport, aggregate analysis of statistics such as OD matrix estimation, ridership, service evaluation and travel time distributions, and the statistical distributions thereof. Moreover, unsupervised learning methods other than clustering that could also be used to conduct pattern knowledge discovery from smart card mobility data are also considered out-of-scope in this survey.

All of the studies included in our analysis were published in the last decade, the vast majority (84%) of which in the past five years. The selected papers were then categorised based on whether they analyse individual mobility patterns or network and urban analytics. Based on the papers reviewed, an additional distinction was made between studies that characterise individual mobility patterns based exclusively on their temporal patterns and studies that also consider spatial features as part of the pattern identification.

In order to address the research questions posed in the previous subsection, we first defined and carried out the search strategy described above. For each individual article included in this review we critically examined and noted the following: study aim, analysis approach (including variables used in performing the clustering), the clustering technique(s) used, any additional variables that were used in the analysis or interpretation that we have identified, the characteristics of the application in terms of geographical context, transport modes included and whether the fare collection system consists of check-in only or also check-out (since it has consequences for whether destinations are observed or inferred) and summarising the key substantive findings. We then synthesised the review results in relation to objectives, methods, applications and results and identify relevant categories of studies and related trends. The following section, Section 2, provides a synthesis of the literature concerned with identifying and characterising travellers’ mobility patterns whereas Section 3 reviews studies that use the latter to characterise individual stations (stop), lines or urban areas. Next, we reflect on the state-of-the-art and point related shortcomings (Section 4) and remaining knowledge gaps, based on which we make suggestions for future research and offer a research agenda (Section 5).

2. Characterising individual mobility patterns

Mobility patterns are characterised by their temporal and spatial features. A careful inspection of the review results reveals that majority of the past work devoted to analysing individual mobility patterns using smart card data has focused on their temporal characteristics, and is reviewed in the following sub-section 2.1. Thereafter, we turn in sub-section 2.2 into reviewing works that have explicitly considered both spatial and temporal aspects of individual’s travel patterns as input to the market segmentation analysis.

2.1. Temporal patterns of individual’s travel

Table 1 provides a summary of the 22 studies analysing temporal travel patterns, including their main aim, the variable(s) which are subject to analysis, the clustering technique employed, additional information that was used in the analysis and in the interpretation of the results, and the key features of the case study application.

Table 1. Summary of studies clustering individuals’ temporal travel patterns.

Study	Aim	Analysis	Clustering technique	Additional analysis and/or interpretation information	Application			Key findings w.r.t patterns identified
					City/Region	Modes	Tap- in (I) / Tap out (O)
Kieu et al., 2015a	Identify users that travel at regular times	Based on the intersection of % of regular OD trips and % of habitual time trips	A-priori segmentation of users (after DBSCAN of stations)	Trip attributes (travel time, mode, transfers), product type	South East Queensland, Australia	BCF	IO	Four segments defined a-priori: “Transit commuters”, “Regular OD passenger”, “Habitual time passenger” and “Irregular passenger”.
Goulet-Langlois et al., 2016	Identify users with similar activity sequence structure	Principal component analysis of multi-week activity sequences	Hierarchical	Socio-demographic information	London, United Kingdom	BMC	I – B; IO- MC	11 segments distinguished based on temporal profiles. The profiles of 40% of frequent users are not characterised by a conventional working pattern; for 9 out of the 11 segments, travel segments are strongly associated with socio-demographic characteristics
Long et al., 2016	Identifying “extreme” users	Percentage and frequency of trips in certain time periods or of certain length	NA	Household survey, inferred home and job locations	Beijing, China	BM	IO (for most trips)	2% of all travellers are labelled “extreme” falling into one of the following cases (in order of prevalence): “Recurring itinerants”, “Night owls”, “Early birds” ad “Tireless itinerants”
Briand et al., 2017	Year-on-year changes	Day-of-the-week and start hour per trip	Two-step Gaussian mixture model	Fare products, spatial entropy of travel destinations	Gatineau, Canada	B	I	Stable shares of (10) temporal profiles over the years
El Mahrsi et al., 2017	Identify groups with similar boarding times	Number of trips per hour-day	Multinomial mixture model	Fare products	Rennes, France	MB	I	13 clusters, morning travel behaviour is more regular and consistent than the evening one
Ghaemi et al., 2017	Reduce high-dimensionality data by transforming the temporal travel representation	Three-dimensional clock-like space projection of hour-day travel	Hierarchical		Gatineau, Canada	B	I	18 clusters distinguished based on travel frequency and typical temporal profile
He et al., 2020	Assessing impacts of different distance metrics	Time-series	Hierarchical		Gatineau, Canada	B	I	A different number of clusters resulting from cross-correlation distance (11 clusters) and dynamic time warping (6) with the former found more suitable
Kieu et al., 2018	Scalability of spatial-behavioural segmentation	Frequency of use, frequency of train use, frequency of transfers, number of unique tap-in and tap-out pairs	Spatial affinity propagation, k-means, Hierarchical	Inferred home location, areas	New South Wales, Australia	BTC+	IO	The proposed Spatial affinity propagation algorithm obtains nearly as good clustering results as the benchmark methods within half the computation time
Manley et al., 2018	Identifying regularities in users’ travel patterns	Time series with binary indications for travel mode and specific services	DBSCAN	Stations	London, United Kingdom	BMC	I – B; IO- M	High travel regularity observed for trips originating from suburbs and destined to central areas
Medina, 2018	Common weekly patterns	Start time and duration of activities	DBSCAN	Household survey	Singapore	BTM	IO	More than 100 clusters of 14-dimension vectors containing information on start time, duration and work or study activities over 7 days. Largest cluster corresponds to 5-weekday workers.
Yang et al., 2018	Predict travel frequency per user based on user segments	Entropy metric related to number of locations visited per time-of-day and weekday-weekend periods	k-means		Shenzhen, China	M	IO	“Regular”, “Variable” and “Irregular” users for which Markov models can predict with high accuracy at the individual level
Cui & Long, 2019	Analysing intra-user regularity of travel patterns for “extreme” and “non-extreme” users	Number of days travelled per week, Travel frequency per day of the week and temporal differences between daily trips	A variant of DBSCAN (denominated OPTICS)		Beijing, China	BM	IO (for most trips)	34 clusters, 4 of which are considered extreme users (“Early birds”, “Night owls”, “Tireless itinerants”, “Recurring itinerants”)
Deschaintres et al., 2019	Intra-user regularity of weekly patterns	Number of trips made on each day of the week	k-means for clustering usage-weeks, Hierarchical for clustering users	Fare products	Montreal, Canada	BM	I	12 user segments with monthly pass users exhibiting greater variability in their usage patterns than annual pass holders
Ji et al., 2019	Determine the socio-economic and job-house relation factors influencing metro commuting patterns	Commuting distance, boarding time to work and boarding time from work	Gaussian mixture model	Household survey	Nanjing, China	M	IO	Three clusters of “classic”, “off-peak” and “long-distance” commuters and nine types of job-housing relations
Viallard et al., 2019	Week-on-week evolution	Number of trips per day of the week	k-means		Gatineau, Canada	B	I	Six clusters with mostly stable patterns but some showing seasonality
Egu & Bonnel, 2020	Measuring inter-user variability over days	Number of days on which both users have travelled	Hierarchical	Fare products	Lyon, France	BTM	I	Intrapersonal variability on weekdays is almost as high as on weekends
Lei et al., 2020	Temporal motifs for travel patterns	Construct a temporal graph of trip sequence, statistical analysis	NA	Land-use	Nanjing, China	M, Bike-sharing	IO (no inter-mode transfers)	11 motifs describe 98% of the data, most are common to metro and bike-sharing trips. The most common motifs are round, chain, single-trip and loop.
Liu & Cheng, 2020	Extract travel patterns, enrich with demographics, illustrate for planning purposes (night tube)	Trip counts per day-of-the-week and start hour	k-means	Inferred home location, demographics	London, United Kingdom	MC	IO	11 clusters that fall into three categories of “Regular peak-time”, “Off-peak noon” and “Randomly evening/ weekend”
Moradi & Trepanier, 2020	Measuring intra-user pattern stability	A binary vector of hour-day travel	Hierarchical (using a semicircle projection)		Gatineau, Canada	B	I	50% of users are found to exercise “high stable” temporal patterns, 40% defined as “mid stable” and 10% “unstable”
Eltved et al., 2021	User groups response to long-term disruptions	Regularity and intensity of travel as well as a weekday/weekend indicator	k-means		Copenhagen, Denmark	MT	IO	8 clusters distinguished based on the frequency of travel on different days-of-the-week and activity type (e.g. “rare weekday”, “occasional leisure”). Commuters reduced most their travel after a disruption.
Liu et al., 2021	Role of age and environments on travel by older users	Entropy metric related to trip frequency per hour	k-means	Personal user information available from card registration	Shizuoka, Japan	BC	IO	Two clusters: low-frequency and high seasonal variability (87% of older adults) and high-frequency and low seasonal variability (13%). The latter tend to be the younger-old (65-74) and live in highly developed areas.
Cats & Ferranti, 2022a	Identify market segments in terms of habitual temporal travel patterns	Number of trips per day-of-the-week and hour-of-the-day combinations	Day of the week analysis: k-means followed by hierarchical; Hourly week pattern analysis: Gaussian Mixture Model	Inferred home location	Stockholm, Sweden	BTMC+	I	Five clusters of “Weekday commuters”, “Lower peaks”, “Late travellers”, “Early birds” and “Flat curve”

Modes: B- Bus; T – Tram; M- Metro; C – Commuter train; + – other modes (i.e. ferry, funiculars).

DBSCAN – Density-Based Scanning Algorithm with Noise; ISODATA – iterative self-organising data analysis technique.

A critical review of the literature reveals an important distinction – yet only seldom made explicit by the studies themselves – between studies focusing on the intra-personal variability of travel patterns versus those concerned with the inter-personal variability of travel patterns. The former aims at classifying individual in terms of the stability of the temporal features of their travel patterns whereas the latter aims at segmenting individuals based on the characteristics of their regular temporal patterns. In other words, intra-personal variability is concerned with how regular one’s travel patterns are. In the context of temporal patterns this can be measured in terms of the extent to which a given traveller boards within the same time periods on given days of the week over a long period of time. In contrast, inter-personal variability pertains to the extent that different travellers exhibit the same travel patterns. In the context of temporal patterns, this may refer to whether different travellers are likely to board in the same time instances over the course of the analysis period. Note that while those categories of studies are not mutually exclusive – one can cluster travellers in terms of their individual travel stability – the distinction pertains to the research focus adopted by the two abovementioned categories.

Studies focusing on intra-personal stability have used various temporal features to analyse travel stability. The latter was investigated in terms of departure times – i.e. a traveller is considered stable if the start times of the trips are consistent over time (Kieu et al., 2015a; Manley et al., 2018) – by the number of trips made on each day of the week – i.e. a traveller is considered stable if the number of trips per day of the week is consistent throughout the analysis period (Deschaintres et al., 2019) – or a combination thereof (Moradi & Trepanier, 2020). Note that the former sets a stricter requirement for being considered a stable traveller than the latter does. A couple of studies have specifically looked into identifying so-called “extreme” travellers defined by the percentage and frequency of trips in certain time periods or of certain duration (Long et al., 2016) or in terms of the number of days travelled per week, travel frequency per day of the week and temporal differences between daily trips (Cui & Long, 2019).

Segmenting individual travellers based on their inter-personal variability can reveal similarities and differences in their temporal travel habits. Several studies have investigated aggregate indicators of travel frequency such as overall frequency of use and frequency of travel characteristics such as travelling by train and performing transfers (Kieu et al., 2018) or number of travel days (Egu & Bonnel, 2020). Since weeks are considered to be the fundamental unit of recurrent travel schedules, the number of trips per day of the week has been often considered an important feature for market segmentation (Viallard et al., 2019), especially in combination with the starting hour (Briand et al., 2017; El Mahrsi et al., 2017; Liu & Cheng, 2020; Cats & Ferranti, 2022a), resulting in an hour-by-hour weekly travel profile per user which is then subject to clustering. User segmentation can then also be used for predicting the travel frequency per user per time-of-day as a conditional probability (Yang et al., 2018). Individual travel diaries allow to specifically focus on the timing of (certain) activities such as boarding times when travelling to and back from work (Ji et al., 2019), the start times as well as duration of activities (Medina, 2018) and the activity sequence structure (Goulet-Langlois et al., 2016; Lei et al., 2020).

Smart card data facilitates the construction of individual-level travel diaries, which in turn enable the investigation of travel patterns amongst specific user groups. Many of the studies reviewed have analysed how the temporal patterns identified vary amongst users with different fare product types using the latter as a contextual post-analysis variable, some of which contain information on different user groups such as school pupils, higher education students, and retired or older users (Kieu et al., 2015a; Briand et al., 2017; El Mahrsi et al., 2017; Deschaintres et al., 2019; Egu & Bonnel, 2020). A single study has specifically focused on a selected user group, namely on the variability of trip frequency per hour by older travellers (Liu et al., 2021). In several studies, personal information available from smart card registrations has also been used, for at least a sample of users included in the analysis (Goulet-Langlois et al., 2016; Liu et al., 2021) or from a matching household survey (Long et al., 2016). However, in most cases no socio-demographic data is available for individual card holders. Notwithstanding, by inferring card-holders place of residence from their travel patterns, one can link users to zonal socio-demographic characteristics available from census data such as ethnicity, employment and income (Liu & Cheng, 2020; Cats & Ferranti, 2022a).

Clustering or segmentation techniques aim at identifying mutually exclusive and collectively exhaustive subsets of the population so as to maximise the intra-group similarities and minimise the inter-group similarities. The definition of the (dis)similarity metric is therefore crucial. In the context of market segmentation, the two most common clustering techniques are $\mathit{k}$-means (see Likas et al., 2003) and agglomerative hierarchical (see Rokach & Maimon, 2005) clustering. The former finds the best partitioning of the dataset given a pre-defined number of clusters, $\mathit{K}$, with the distance metric pertaining to the centre of the cluster. A variant of which, $\mathit{k}$-medoids, selects an actual data point as the centre of each cluster. Agglomerative hierarchical techniques follow a bottom-up approach with iterative merging of groups of observations based on their similarity. While hierarchical approaches are more computationally expensive, they allow for greater variety of distance specifications and their output contains more information on relations between potential clusters and does not require the a-priori specification of the number of clusters. An additional technique used in several studies for clustering users based on the similarity between subsequent trips is DBSCAN (Khan et al., 2014) which relies on user-specified parameter values. A model-based approach such as Gaussian mixture model offer a latent profile assignment with probabilistic assignment of members to clusters which is especially attractive in the case of continuous variables (e.g. Briand et al., 2017).

A variety of clustering techniques have been employed in the analysis of temporal travel patterns. The most commonly used methods are: (i) K-means (Kieu et al., 2018; Yang et al., 2018; Deschaintres et al., 2019; Viallard et al., 2019; Liu & Cheng, 2020; Liu et al., 2021; Eltved et al., 2021), (ii) DBSCAN (Medina, 2018; Manley et al., 2018; Cui & Long, 2019), (iii) hierarchical clustering (Goulet-Langlois et al., 2016; Ghaemi et al., 2017; He et al., 2020; Kieu et al., 2018; Deschaintres et al., 2019; Egu & Bonnel, 2020; Moradi & Trepanier, 2020; Cats & Ferranti, 2022a) and (iv) multinomial or Gaussian mixture models (Briand et al., 2017; El Mahrsi et al., 2017; Ji et al., 2019; Cats & Ferranti, 2022a). A single study applied a spatial affinity propagation technique (Kieu et al., 2018). Several studies have specifically focused on improving the projection of temporal distances to facilitate its hierarchical clustering (Ghaemi et al., 2017), on comparing the performance of alternative distance metrics when clustering users based on time-series data (He et al., 2020) or comparing alternative clustering techniques (Kieu et al., 2018).

Case study applications analysed smart card data from cities in Australia (Brisbane, Sydney), Canada (Gatineau, Montreal), China (Beijing, Nanjing, Shenzhen), Denmark (Copenhagen), France (Lyon, Rennes), Japan (Shizuoka), Singapore, Spain (Tarragona), Sweden (Stockholm) and United Kingdom (London). Most applications have considered metro systems, bus systems or a combination thereof, with few studies extending to other urban and suburban modes of public transport. Smart card data from most metro systems worldwide and bus systems from systems in East Asia analysed by studies included in this review contain both tap-in and tap-out data. Note that for most approaches taken in the analysis of temporal travel patterns no information regarding travel destination is per-se needed since boarding time (available from tap-in transactions) is the prime variable of interest used for constructing travel profiles that are then subject to clustering.

Few of the studies examining temporal travel patterns have also considered spatial elements such as the share of regular OD trips (Kieu et al., 2015a) or information on travel modes and specific services (Manley et al., 2018). In addition, a number of studies have considered spatial attributes such as relating those to stations (Manley et al., 2018), inferred home location (Kieu et al., 2018; Liu & Cheng, 2020; Cats & Ferranti, 2022a), inferred home and job locations (Long et al., 2016), land use information (Lei et al., 2020), diversity of travel destinations as measured by means of spatial entropy (Briand et al., 2017) as additional variables used for enriching the analysis and interpretation of the results. Notably, these variables range from incorporating information from other datasets such as land use registration or census survey to analysing the share of each identified user segment for different spatial resolutions (e.g. stations, zones).

A synthesis of the key findings reveals that patterns are characterised in relation to their travel frequency, days-of-the-week and times-of-the-day travel with results consistently distinguishing between regular/commuting travellers and irregular/occasional travellers categories with several clusters therein, especially in relation to a range of irregular and relatively uncommon patterns (e.g. Goulet-Langlois et al., 2016; El Mahrsi et al., 2017; Yang et al., 2018; Liu & Cheng, 2020; Cats & Ferranti, 2022a). The regular/commuting travellers typically constitute a large and more homogenous minority of card-holders yet they perform the vast majority of trips.

2.2. Spatiotemporal patterns of individual users’ travel

Table 2 provides a summary of user segmentation studies that have identified spatiotemporal travel patterns. Similarly to the main divide in the analysis of temporal travel patterns, also studies focusing on both spatial and temporal characteristics have either looked into intra-personal or inter-personal variability. The former considered in this context thus not only the extent to which users travel at the same times over the course of the analysis period (be it defined in terms of frequency per week, days of the week and/or time of the day) but also the locations between which they have travelled.

Table 2. Summary of studies clustering individual’s spatial-temporal travel patterns.

Study	Aim	Analysis	Clustering technique	Additional analysis and/or interpretation information	Application			Key findings w.r.t patterns identified
					City/ Region	Modes	Tap- in (I) / Tap out (O)
Ma et al., 2013	Distinguish between users based on the regularity of spatiotemporal characteristics	Spatial and temporal features such as number of travel days, number of similar first boarding times and similar stations and route sequences	DBSCAN per user and k-means to cluster users		Beijing, China	BM	IO (for most trips)	Five clusters varying from very low to very high travel regularity
Goulet-Langlois et al., 2018	Measuring regularity of users’ travel patterns in terms of frequency and sequence	Sequence of activity locations and respective activity durations per day	NA (regular and irregular users are identified based on percentile values of an entropy metric)		London, United Kingdom	BMC	I – B; IO- M	Much of the regularity and predictability of travel patterns is captured by the order of travel events (predictions are three times more accurate)
Ma et al., 2017	Identifying commuter/ regular users	Number of days travelled, number of occurrences for most frequent start location, destination location and respective boarding times and routes	DBSCAN followed by ISODATA (an unsupervised learning method)	Online survey, areas between ring-roads	Beijing, China	BM	IO (for most trips)	“Absolute commuters”, “Average commuters” and “Noncommuters” are distinguished and can be retrieved with high accuracy based on a survey conducted
Wang, Zhang et al., 2020	Identify irregular (undirected) users	Weighted sum of travel frequency at visited stations and time periods	k-means for temporal patterns followed by DBSCAN for detecting irregular travellers	Stations	Beijing, China	M	IO	9.5% of all activity sequences are identified as “irregular” which are then clustered into 13 groups in relation to time spent travelling and unusual travel speed
Gutierrez et al., 2020	Latent profile analysis of tourists travel patterns	Variables related to number of active days, number of transactions, average group size, number of stations, municipalities and zones visited, main stations visited, number of routes used, and share of transactions at main stations and areas.	Gaussian mixture model	Stations	Tarragona, Spain	B	I	Five visitor classes characterised by party size and, the duration and intensity of their use
He et al., 2021	Integrate spatial- and temporal aspects of daily behaviour	Dynamic time warping distance of station-sequences in daily profile	Hierarchical clustering (with sampling)		Gatineau, Canada	B	I	Ten clusters of users in terms of daily travel trajectory that tend to have distinctive boarding areas
Kaewkluengklom et al., 2021	Year-on-year changes in cluster membership	Distribution of trips over four daily time periods, top three most used ODs and most used stations	k-means	Card type, land-use type and density, population density	Shizuoka, Japan	BC	IO	Ten clusters with two main profiles – one for regular commuters and one for holiday and specific activities users. Regular commuters show stable patterns and clusters’ size remained stable
Pieroni et al., 2021	Identify the travel patterns of low-income workers	Nine variables pertaining to the four categories of temporal patterns and variability, spatial patterns and variability, activity pattern and variability, and socio-economic.	k-means	Land-use types	Sao Paulo, Brazil	BMC	I	Eight user clusters, two of which show commuting patterns which are characteristic of low-paid employment (early first transaction, longer elapsed time between transactions, high usage frequency), also confirmed by relating those to origin/destination land-use and socio-demographic characteristics
Zhang, Yang et al., 2021	Examine variations in activity spaces of vulnerable users’ groups	Indicators related to travel frequency, travel distance, radius and shape index of activity space, and place and time entropy to reflect diversity of activities per month	k-means	Population groups based on card type (low-income, disabled and elderly)	Wuhu, China	B	I	Low-income users, disabled users and users residing in suburbs travel more frequently, longer distances and have a larger activity range than the population average. The elderly users travel more frequently, similar distances to population average and have a smaller activity range.
Faroqui et al., 2022	Identify targeted groups of users based on their activities and trips for advertising purposes	Similarity in terms of location, time and type of activity	Agglomerative hierarchical technique (Ward method)	Activity location type (station, vehicle, billboard) and purpose of ads (work, education, shopping)	South East Queensland, Australia	BCF	IO	91 activity-trip groups defined as users with similar type, time and location of activity and spatial-temporal features of their trip
Cats & Ferranti, 2022b	Identify the extent to which users visit different and distinctive parts of the network	Share of user-trips destined to each zone, and the probability that a user visits a zone in relation to the overall visiting volume of this zone	K-means (for visiting frequency) and Gaussian Mixture model (for spatial extent)	Inferred home zone, and respective zonal income, social index and share of foreign background	Stockholm, Sweden	BTMC+	I	Three zonal visiting frequency profiles of “Local”, “Commuter” and “Explorer” with varying home-zones and socio-demographics. 15 visiting spatial extent profiles with diverse geographical orientations.

In an attempt to identify commuters, studies focusing on intra-personal variability have considered variables such as the number of active travel days and the similarity of first boarding times, stations and route sequences (Ma et al., 2013) or the prevalence of most frequent origin-destination and the respective boarding times and routes (Ma et al., 2017). A broader perspective, not limited to commuting and most commonly performed trips, considers the distribution of trips over daily time periods and the most often used ODs and stations (Kaewkluengklom et al., 2021), or adds greater detail in the form of the overall sequence of activity locations and their respective durations per day (Goulet-Langlois et al., 2018; He et al., 2021). A similar approach served the opposite goal in identifying irregular (undirected, potentially suspicious) users based on travel frequency per time window and station (Wang, Zhang et al., 2020).

Studies focusing on inter-personal variability group users based on the characteristics of their regular spatiotemporal travel patterns (rather than the extent to which those are stable). Four studies have been identified in this category, three of which focused on special user groups. Gutierrez et al. (2020) examined the combination of an array of temporal (travel frequency and number of travel days) and spatial (number of locations visited and routes used, prominence and identity of most visited locations) characteristics of tourists’ travel patterns. Wang, Zhang et al. (2020) investigated the activity spaces of vulnerable users’ groups and employed to this end indicators related to both temporal (frequency, time entropy) and spatial (distance, radius and shape of activity space, place entropy) travel characteristics. Pieroni et al. (2021) focused on low-income workers in the context of a metropolis with large income gaps and contrasts their travel patterns with those of high-income ones. The analysis conducted by Cats and Ferranti (2022b) is the only one within this category which was not limited to any specific user group. Their analysis quantifies the spatial extent of locations visited, i.e. trip destinations, by users within the course of a period. A snowball search identified one additional study which regressed the extent to which users tend to travel using the same route against user, OD and trip characteristics, albeit without seeking to identify patterns (Kim et al., 2017).

Two techniques are dominant when integrating spatial information into the clustering of individual users, k-means (Ma et al., 2013; Yang et al., 2018; Kaewkluengklom et al., 2021; Wang, Zhang et al., 2020; Zhang, Yang et al., 2021; Pieroni et al., 2021; Cats & Ferranti, 2022b) and DBSCAN (Ma et al., 2013; Ma et al., 2017; Wang, Zhang et al., 2020). Some exceptions exist for studies using a diverse array of travel features which use a Gaussian mixture model (Gutierrez et al., 2020; Cats & Ferranti, 2022b) or hierarchical clustering (He et al., 2021), as well as a study identifying irregular users based on percentile cut-off of an entropy-based metric (Goulet-Langlois et al., 2018).

Of the eleven studies identified which analysed the spatiotemporal characteristics of individual mobility patterns using smart card data, three of which have been conducted for the case of Beijing and the remaining, with one study each, for Brisbane, Gatineau, London, Sao Paulo, Shizuoka, Stockholm, Tarragona and Wuhu.

A special class of studies – not concerned with the identification of user segments, yet concerned with their individual travel patterns – is the one focusing on physical encounters and thereby contact networks and their potential consequences within in the context of virus spreading. All these works construct a contact network based on the inferred travel trajectories. Sun et al. (2013) analysed the statistical properties of the topological indicators – including the clustering coefficient – of the resulting physical encountering network for bus users in Singapore. Liu et al. (2020) searched for the co-existence of users at stations and on-board vehicles using DBSCAN. Their application considered the metro networks of Shenzhen and they analyse the frequency and duration of encounters among pairs of users. Qian et al. (2021) analysed the dynamics of virus spreading in contact networks and the effectiveness of control strategies, namely vaccination and quarantine, using contract networks constructed from the metro systems of Guangzhou, Shanghai and Shenzhen in China.

A synthesis of the findings reported suggests that there are few prevailing patterns in relation to spatial-temporal travel regularity which typically pertain to a different degree of commuting vs. occasional travel (e.g. Ma et al., 2013; Ma et al., 2017; Gutierrez et al., 2020; Kaewkluengklom et al., 2021). Moreover, there is evidence to suggest that those can be reasonably well predicted (Goulet-Langlois et al., 2018). However, once more information on spatial features of travel characteristics is considered, a great diversity of travel patterns is observed (e.g. Faroqui et al., 2022; Cats & Ferranti, 2022b).

3. Spatiotemporal travel patterns for network and urban analytics

In the previous section we reviewed the literature on clustering public transport users based on the temporal or spatial-temporal characteristics of their individual travel patterns. Another, related, body of literature has sought to identify and characterise clusters of (public transport) network elements – most often stations but in a few cases service lines – based on the spatial-temporal characteristics of their respective users, as well as the urban area in which the networks are embedded.

Table 3 presents a summary of studies that performed network and urban analytics based on spatiotemporal travel patterns derived from smart card data. The objectives of the 24 studies included in this review vary greatly. Pioneering studies in the field sought to identify key activity centres using users’ flow data (Roth et al., 2011; Kim et al., 2014; Cats et al., 2015). Those have been followed by a series of studies that investigated the relation between the results of the station clustering and land-use patterns and places of interest in proximity to these stations or characteristics of the respective zones (Kim et al., 2017; Zhang et al., 2018; Zhao et al., 2019; Gan et al., 2020; Kim, 2020; Zhuang et al., 2020). Recent studies have extended this to the analysis of how urban structure has evolved over time (Wang et al., 2021; Zhang, Marshall et al., 2021). Another stream of works has focused on the generation of origin-destination matrices by aggregating stations into demand zones (Kieu et al., 2015b; Luo et al., 2017), characterising the demand patterns associated with individual stations (Zhong et al., 2015; El Mahrsi et al., 2017; Wang et al., 2017; Li et al., 2020; Zhou et al., 2022; Park et al., 2022), and clustering stations (Cats & Ferranti, 2022c; Chan Chun et al., 2023) or lines (Wang, Luo et al. 2020; Yap et al., 2019) that exhibit homogenous activity patterns.

Table 3. Summary of studies clustering network elements based on related travel patterns.

Study	Aim	Analysis	Clustering technique	Additional analysis and/or interpretation information	Application			Key findings w.r.t patterns identified
					City/Region	Modes	Tap- in (I) / Tap out (O)
Roth et al., 2011	Identify the polycentric structure of activity centers	Descending order to inflows and bundling stations in proximity	Hierarchical (own variant)	Travel distance	London, United Kingdom	M	IO	Ten clusters corresponding to different geographical parts of the city with few centers attracting most flows and revealing a polycentric structure
Kim et al., 2014	Identify zones and related flows	Average frequency of flows	Hierarchical (own variant)		Seoul, South Korea	M	IO	1745 movement patterns between zones related to both origin and destination areas that obtained best coverage and accuracy metrics
Cats et al., 2015	Identifying and classifying activity centers	Temporal profile of metrics related to boarding and alighting flows	Two-stage hierarchical for identification and classification of geographical clusters		Stockholm, Sweden	BTMC	I	Identifying 17 activity centers which are then partitioned into 6-8 classes for different boarding and alighting flow relations. Limited evidence for the realisation of Stockholm’s policy to become more polycentric.
Kieu et al., 2015b	Computational efficiency of travel pattern clustering	Based on OD pairs	Modified DBSCAN		South East Queensland, Australia	BC+	IO	Proposed algorithm obtains the same clustering results as the classical DBSCAN in a less than 1% of the computation time
Zhong et al., 2015	Variability of relations between stations over time	Group stations with high inter-flows	Community detection		Singapore	BM	IO	The number of clusters varies between 7 and 9 for different days of the week and some stations exhibit different patterns (i.e. belong to different clusters) on different days of the week
El Mahrsi et al., 2017	Identifying stations’ roles based on temporal profile	Boarding flows per hour and day per station	Poisson mixture model		Rennes, France	MB	I	14 clusters that belong to two main categories: mostly balanced with some peaks during rush hours or unbalanced between morning and afternoon
Kim et al., 2017	Relation between travel patterns and station area	Boarding and alighting flows per hour and day of the week	k-means	Land-use and socio-demographic variables	Seoul, South Korea	M	IO	Five clusters described as commercial (business), commercial (entertainment), mixed (commercial), mix (residential) and residential.
Luo et al., 2017	Identify clusters of stations with similar travel patterns for data-driven OD generation	Ratios between intra-cluster flows and inter-cluster flows and distances	k-means	Time of day and weekdays/ weekends	The Hague, the Netherlands	BT	IO	12 geographically contiguous clusters with general patterns (number and composition of clusters) remaining stable over different time periods
Wang et al., 2017	Identify functions of stations using static geographical information	Origin and destination stations	Affinity propagation	Points of Interest (POI) characteristics; time of day and weekdays/ weekends	Shanghai, China	M	IO	Ten clusters such as “High-tech development regions” and “Emerging residential regions” when combining static POI characteristics and dynamic mobility data
Zhang et al., 2018	Estimate the impact of zonal variables on the spatial distribution of demand and differences between public transport and taxi	Average daily trips	Community detection	Population density and average income per zone	Singapore	BM, Taxi	IO	Communities identified based on smart card data differ for weekdays and weekends whereas for taxi data no such difference is observed and the communities are spatially more coherent and aggregate
Yap et al., 2019	Identify hubs and determining line bundles to be prioritised for transfer synchronization	Constructing a graph representing transfer flows between neighbouring stations	DBSCAN for hub identification followed by modularity-based community detection for line bundles determination	Time periods	The Hague, the Netherlands	BT	IO	Identifying 23 clusters for 696 transfer locations, more than 70% of the transfers take place at 6 of which (“hubs”). For those hubs “line bundles” are identified for transfer synchronization.
Zhao et al., 2019	Correlation analysis between clustering results and surrounding land uses	Normalized hourly ridership per station	k-medoids	Land uses	Nanjing, China	M	IO	Five clusters focused on city centre and sub-centers, border between core and suburbs, main urban area and university towns, industrial parks, and major long-distance railway hubs
Gan et al., 2020	Explaining boarding and alighting flows profile with local environment variables	Normalized boarding and alighting flows per hour	k-means	Land-uses	Nanjing, China	M	IO	7 clusters distinguished based on land-use and main trip generators and attractors in proximity
Kim, 2020	Enhance urban structure analysis with spatial mobility patterns (in addition to temporal)	Similarity between the distribution of boarding flows and alighting flow probabilities per time period	Spectral clustering followed by hierarchical clustering	Land-use, residential and employment density	Seoul, South Korea	MB	IO	Dividing the territory into 25 clusters with high internal homogeneity in relation to land-use mix, number of residents and jobs
Li et al., 2020	Identify important features for classifying stations	PCA of eight indicators that relate to shares of users in different periods and statistics of the time series of users counts	Own 2-step heuristic		Beijing, China	M	IO	Six clusters ranging from “Residential and office mixed type but partial living type” (double-peak distributing on working days) to “Transportation hub and tourism commercial type” (no obvious working day peak)
Wang, Luo et al., 2020	Identify the hierarchy amongst lines	Representing users flows using a dual graph (C-space)	Community detection	Time periods	The Hague, the Netherlands	BT	IO	Four communities of lines in terms of network hierarchy. Tram lines tend to rank higher than bus lines, especially those traversing connecting rail stations.
Zhuang et al., 2020	Relate stations to each other and location type	Boarding flow per time of day period and day	Affinity propagation	Points of Interest (POI) characteristics	Singapore	BTM	IO	Nine clusters of POI location semantics with stations groups labels such as “Under-development areas”, “Old residential areas” and “Emerging commercial areas”
Wang et al., 2021	Identifying functional areas and their changes over years	Frequency of visits per station per time period	Gaussian mixture model		Beijing, China	M	IO	Five clusters denoted as “Multimodal interchange hubs and leisure”, “ Residential”, “Employment”, “Mixed but mainly residential” and “Mixed employment and residential”, evolution happens in increments through mixed phases
Zhang, Marshall et al., 2021	Explaining changes in movements over years using station area variables	Weekday peak time flows	Community detection	Socio-economic data at neighbourhood-level (population and job density, median house price, planning policy)	London, United Kingdom	M	IO	London became more polycentric and compact between 2013 and 2017. Five communities (geographically contiguous) are identified in both periods
Cats & Ferranti, 2022c	Identify and classify urban areas based on how attractive they are as activity centers	Share of visits by non-residents (based on home-zone inference) from all visits for stations included in a clustered area per day-of-the-week and hour-of-the-day combination	k-means	Time periods	Stockholm, Sweden	BTMC+	I	Four key clusters based on their zonal visiting profile labelled “Primary attraction”, “Secondary attraction”, “Low intensity destinations” and “Residential destinations” with a clear centre-periphery order
Park et al., 2022	Forecast daily pattern of boarding flows	Boarding flows per time window	Discriminative functional mixture model (fun FEM)		Seoul, South Korea	M	IO	Six clusters based on the temporal profile of alighting flows, resonates with typical land-use and activities in proximity
Yong et al., 2021	Identifying employment, residential and balanced station types	Job-housing ratio and commute properties	k-means	Residential and employment density	Chongqing, China	M	IO	“Residential”, “Employment” and “Balanced” clusters based on boarding and alighting flow ratios throughout the day
Zhou et al., 2022	Topological and boarding and alighting flow characteristics	Boarding and alighting flows as well as degree and betweenness centrality	k-means		Beijing, China	M	IO	Six clusters in relation to absolute flow levels throughout the day
Chan Chun et al., 2023	Classifying stations in relation to their urban function	Numbers of boarding and alighting flows per hour	k-means and a principal component analysis (PCA)	Land use, population, user travel times	Seoul, South Korea	M	IO	Four groups: “Main commercial”, “Common commercial”, “Residential” and “The others”. The former consists of four business areas of varying spheres of influence (strength).

With the exception of the latter group focusing on service lines which is aimed at supporting service planning, all other studies have used individual stations or set of stations in proximity (i.e. areas) as the unit of analysis subject to clustering. All of the abovementioned studies have utilised origin-destination flow information derived from smart card data in their analysis with the exception of one study relying on boarding volumes only (El Mahrsi et al., 2017) and two studies considering boarding and alighting flows (Cats et al., 2015; Gan et al., 2020), presumably due to data availability limitations.

Given the dynamic character of network and urban centre characteristics, it is not surprising that the vast majority of studies have considered the temporal properties of users’ flow distribution in their analysis. Most studies have integrated information on the temporal characteristics of the flows analysed as part of the clustering process (Cats et al., 2015; El Mahrsi et al., 2017; Kim et al., 2017; Zhao et al., 2019; Gan et al., 2020; Kim, 2020; Li et al., 2020; Zhuang et al., 2020; Wang et al., 2021; Zhang, Marshall et al., 2021; Cats & Ferranti, 2022c; Chan Chun et al., 2023). In addition, several studies have investigated how the results of their clustering vary for different time periods (Luo et al., 2017; Wang et al., 2017; Yap et al., 2019; Wang, Luo et al. 2020).

Methods deployed for the clustering of stations (or sets of stations) and lines include k-means or variants thereof (Kim et al., 2017; Luo et al., 2017; Zhao et al., 2019; Gan et al., 2020; Yong et al., 2021; Cats & Ferranti, 2022c; Zhou et al., 2022; Chan Chun et al., 2023), variants of hierarchical methods (Roth et al., 2011; Kim et al., 2014; Cats et al., 2015; Kim, 2020), DBSCAN (Kieu et al., 2015b; Yap et al., 2019), Poisson or Gaussian mixture models (El Mahrsi et al., 2017; Wang et al., 2021), affinity propagation (Wang et al., 2017; Zhuang et al., 2020) and a discriminative functional mixture model (Park et al., 2022). In addition, a number of studies adopted community detection techniques (Zhong et al., 2015; Yap et al., 2019, Zhang et al., 2018; Wang, Luo et al., 2020; Zhang, Marshall et al., 2021). The latter – which have not been adopted by any of the studies performing user segmentation reported in sections 2.1 and 2.2 – is a graph-based technique which partitions a network into communities by grouping together nodes (i.e. typically corresponding to stations) that are densely connected with those belonging to different communities being only sparsely connected, where the strength or distance of a connection is indicated by the selected link labelling.

Largely reflecting the geographical distribution of the research groups that have conducted related research, studies have analysed urban and system structure for cities from Australia (Brisbane), China (Beijing, Nanjing, Shanghai, Chongqing), France (Rennes), the Netherlands (The Hague), Singapore, South Korea (Seoul), Sweden (Stockholm) and the United Kingdom (London). The majority of studies have utilised (only) metro data whereas several studies included all public transport modes present in the case study area to best reflect the underlying travel demand patterns, including in areas which might be underserved by the metro. A key distinction can be made between studies that cluster stations into geographically contiguous clusters (e.g. Roth et al., 2011; Luo et al., 2017; Cats & Ferranti, 2022c) and those that assign each station into one of the clusters independently of its neighbouring stations (e.g. El Mahrsi et al., 2017; Zhuang et al., 2020). The former is more suitable for most travel demand planning and management purposes. In both cases the main distinction is between areas that serve commuting/business/commercial trips and residential areas with reverse morning vs. afternoon patterns (e.g. Cats et al., 2015; El Mahrsi et al., 2017; Kim et al., 2017; Li et al., 2020; Wang et al., 2021).

Given the scope of this review, the discussion above and Table 3 are limited to studies that identify patterns in how user flows differ across the system for the purpose of either network or urban analytics. The bibliometric search yielded several adjacent studies that are either grounded in complex system theory or in urban planning. Studies within the realm of the former investigate statistical properties of travel flows across the network, such as the stability of how travel patterns vary over days for the networks of London, Singapore and Beijing (Zhong et al., 2016) or calculated several measures of urban diversity based on the temporal pattern of flows per spatial units (Sulis et al., 2018). Examples of the latter include the analysis of year-on-year change in the job to worker ratio per station in the Beijing metro network (Huang et al., 2019) and the identification of employment areas in Beijing and the evolution thereof based on temporal characteristics of boarding and alighting flows (Huang et al., 2021). Note that these studies, unlike the studies included in Table 3, use descriptive statistics to observe spatial variations.

4. Discussion

As is evident from the synthesis of the literature provided in sections 2 and 3, there is an increasing body of research that has either characterised user travel patterns or performed network and urban analytics based on the associated travel patterns. In particular, several on-going trends in this research domain can be identified:

• Specific user groups. There is increasing interest in investigating travel patterns for particular target segments which exhibit distinctive travel patterns ranging from tourists (Gutierrez et al., 2020) and older travellers (Liu et al., 2021) to lower income individuals (Pieroni et al., 2021).

• Enriching smart card data with socio-demographic data. This has been accomplished by means of including personal information available from card registrations (Goulet-Langlois et al., 2016; Liu et al., 2021) or a matching household survey (Long et al., 2016). Alternatively, zonal socio-demographic data was linked by inferring the home-zone location of card holders (Liu & Cheng, 2020; Cats & Ferranti, 2022a).

• Evolution of travel patterns. The availability of smart card data over the course of a substantial time period allows for the analysis of changes in observed travel patterns in terms of identified market segments (Briand et al., 2017; Viallard et al., 2019), station clusters (Huang et al., 2021) or underlying urban characteristics (Zhang, Marshall et al., 2021).

When examining the literature devoted to user segmentations, it is evident that the number of studies that consider the spatial characteristics of travel patterns in the identification of user segments is comparatively small compared with the body of literature devoted to temporal aspects of user travel patterns using smart card data. Furthermore, all of the relevant studies identified in this review, except for the notable exception of Cats and Ferranti (2022b), have analysed spatial features in conjunction with temporal ones, rather than analysing spatial aspects in isolation. In line with the scope of this review, studies that analysed the spatiotemporal characteristics of ridership flows or origin-destination matrices are not included in this analysis.

The limited research effort devoted to the consideration of spatial features as part of user segmentation arguably stems from its greater complexity as compared to temporal aspects. The latter can easily be discretized and are universal and therefore the transferability of the associated feature definition and clustering techniques is fairly straightforward. In contrast, spatial features are subject to local variations and are often labelled, their discretization is not trivial and the findings cannot be easily made transferable to other contexts. The main challenge associated with studies that are concerned with clustering – either users or network elements – based on spatial features is how to incorporate those in a generic way as opposed to tailored to local features. It is therefore advisable to refer to topological characteristics which can be transferable to other contexts (e.g. centrality indicators of stations or diversity of locations visited) rather than references/labels of specific locations.

In our review of the literature it also became apparent that certain important aspects related to the underlying data used are under-discussed, hampering transparency and cross-comparison, and consequently hamper the advancement of this research field. In the following we discuss three such aspects:

• Time span covered by the smart card dataset. The time period for which smart card data has been analysed varies greatly among studies, measured in days or years. Time periods vary from a single day (e.g. Faroqui et al., 2022), one (e.g. Cui & Long, 2019) or several weeks (e.g. Goulet-Langlois et al., 2016; El Mahrsi et al., 2017; Ghaemi et al., 2017; Gan et al., 2020) to several months (e.g. Egu & Bonnel, 2020), about a year (e.g. Deschaintres et al., 2019; Cats & Ferranti, 2022a, 2022b, 2022c) or up to several years (e.g. Goulet-Langlois et al., 2018). Several authors commented on the importance of using a sufficiently long period in order to extract meaningful patterns. For example, Goulet-Langlois et al. (2016) assert that it is important to consider several weeks in the analysis because of changes in behaviour from one week to the other. Cui and Long (2019) argue that a week-long analysis is sufficient even for their comparison over two non-consecutive years, whereas Briand et al. (2017) used one month data for each of five consecutive years in their pattern year-on-year changes analysis. There is thus no consensus as to how long of a period should be analysis. This calls for the development of standards and criteria to ensure that patterns reported are based on data collected over a sufficiently long period for a given analysis/application.

• Assumed relation between cards and individual travellers. The identification of user segments based on smart card movements assumes that cards correspond to individual users. In many transit systems worldwide smart cards may either by anonymous or personal with the latter allowing for more complicated subscriptions and concessions. Anonymous cards may be used by multiple individuals thereby violating the underlying assumptions employed for attributing the movements associated with a card to a single individual. This key assumption is however very seldom made explicit in the vast majority of the studies reviewed. A notable exception is made by Eltved et al. (2021) who explicitly limits their analysis of to the impacts of long-term disruptions to the set of personal smart card data in Copenhagen to allow for the disaggregate analysis of behavioural responses for which having panel data is a pre-requisite. In addition, studies focusing on specific user groups who are eligible for concessions such as school pupils, higher education students, and retired or older users (Kieu et al., 2015a; Briand et al., 2017; El Mahrsi et al., 2017; Deschaintres et al., 2019 and Egu & Bonnel, 2020). have also limited their analysis to personal cards.

• Smart card data as a good proxy of travel demand. For studies investigating public transport demand, i.e. individual mobility patterns, it is relevant to know what share of all public transport trips are registered in the form of smart card data. While smart card typically form the main mode of payment, it is seldom the only mean of fare collection. Transit systems worldwide offer a variety of alternative modes of payment including single-use (often paper) tickets and tickets purchased via a smart phone application. Datasets may or may not include transactions made by other means such as debit cards and the scanning of QR codes. The vast majority of studies reviewed do not comment on these aspects which are relevant for appreciating the completeness (and possibly representativeness as some groups might be over- or under-represented in some modes of payment, e.g. tourists). Few notable exceptions include El Mahrsi et al. (2017) who reported that 80% of all trips are registered in the smart card dataset in the case of Rennes, France and Egu and Bonnel (2020) who reported a 75% rate for the case of Lyon, France. Another related aspect which is seldom discussed by studies analysing travel patterns using smart card data pertains to the share of (public transport) trips captured. For studies focusing on urban analytics this is a crucial piece of information.

We therefore recommend that future studies include the following information in their analysis: (i) comment on why the timespan chosen is considered adequate for the relevant application; (ii) make an explicit statement about the type of cards included in the analysis (personal/anonymised) or the composition thereof, comment on potential implications for their clustering analysis, and; (iii) report the share of public transport trips included in the dataset used for analysis as well as the modal share for the analysis area in the case of an urban analytics study. Following such reporting standards will improve the soundness of the reporting and provide context when comparing findings associated with different dataset/systems.

5. Research agenda

Based on our critical review of the state-of-the-art we identify four knowledge gaps and outline four related themes for further research. The first two themes correspond the consideration of spatial features and the availability of panel data, following up on aspects mentioned in Section 4. The latter two themes refer to moving from understanding to predicting and connecting findings to (planning and operations) applications.

• Enhanced geographical characterisation of users’ travel patterns. The analysis of mobility patterns and the related user clustering and market segmentation have insofar largely focused on their temporal features (section 2.1) or the stability and variability of travel destinations (section 2.2). However, there is lack of knowledge, with the notable exception of the work of Cats and Ferranti (2022b), on the geographical properties of individual travellers by characterising spatial features of locations visited based on disaggregate longitudinal data. For the latter to be made available it is essential to comply with rigid privacy protocols due to the sensitivity of sequential (panel) geo-location data. Moreover, the aforementioned trend of enriching smart card data with socio-demographic data has been observed only for temporal user clustering but is yet to enter studies focusing on spatial clustering. The analysis thereof will enable identifying the extent to which different user groups visit different parts of the urban area. An aggregate analysis of the latter will allow unravelling the composition of travellers visiting various zones and the extent to which various destinations are visited by people with different backgrounds. Community detection techniques may be apply for user segmentation with the goal of identifying groups of travellers that share similar geographical features of their mobility patterns.

In the analysis of spatial characteristics it is advised, as mentioned in Section 4, to extract features that can be generalised to other contexts such as density, diversity and geographical scope of destinations visited rather than focus on labelled ones that pertain to local features (e.g. location names). Given the significant role that features of the local urban environment are likely to play in determining the spatial characteristics of users’ mobility, there is a need to gain more knowledge on how those are manifested in diverse settings. While these variables vary also within any given city, they may be confounded and the range of values might be limited. A cross-city comparison will allow concluding on relevant determinants such as urban form, density, availability and quality of alternative modes of transport, land-use and demographics.

• From demand analytics towards behavioural analytics. While the identification and characterisation of human mobility patterns offers remarkable demand analytics insights it does not advance our knowledge on the underlying determinants of user’s behaviour. A separate large body of research is devoted to the estimation of travel choice determinants from smart card data, primarily limited to the estimation of route choice models and the impact of crowding therein (e.g. Hörcher et al., 2017; Yap & Cats, 2021). There is great potential in marrying these two research streams so as to support the development of behavioural models for users with distinctive mobility patterns, for example by means of estimating latent class models that are informed by the user clustering results.

Obtaining behavioural insights will also be stimulated by segmenting mobility patterns based on a combination of travel features that provide additional information on users experience. A couple of recent examples making advancements in this direction include the identification of central users for virus spreading by means of the simultaneous consideration of the distance travelled, the number of fellow travellers one has been exposed to and the degree of exploration (El Shoghri et al., 2020) and the neural network approach taken for the joint consideration of travellers’ departure and arrival time, travel location, travel distance and point of interest at the destination (Li et al., 2021).

• Predictions of travel patterns. The availability of system-wide demand data opens avenues for the development of schemes aimed at the prediction of travel patterns. Such predictions can facilitate service design adjustments as well as supply- and demand-management measures to better cater or steer anticipated flows. A couple of studies included in this review already do so specifically for travel frequency per user (Yang et al., 2018) and the daily pattern of boarding flows at stations (Park et al., 2022).

The future development of prediction schemes can be structured along two axes: short-term vs. long-term and disaggregate vs. aggregate. The former distinction pertains to the time horizon for which the prediction is made, whether it concerns within-day downstream conditions measured in minutes to hours or considering a timespan of days or even months. Short-term predictions pose requirements on computational efforts and possibly also on the real-time availability of smart card data whereas long-term predictions can rely on offline applications. Several recently deployed fare collection systems enable the real-time availability of smart card data, albeit those are still far from common practice. The distinction between disaggregate vs. aggregate pertains, respectively, to whether the unit of analysis that is subject to prediction relates to trip characteristics of each user or to the travel pattern characteristics which are the outcome of collective dynamics, such as boarding, alighting and on-board flows.

All four combinations of short- and long-term, disaggregate and aggregate predictions are relevant as future research directions and can facilitate respective applications and interventions by service providers. Several pioneering studies proposed a logistic regression model for predicting the next trip to be performed by a user (Zhao et al., 2018) and an elasticity model for short-term aggregate predictions (Van Oort et al., 2015).

• Decision support for service planning and policy evaluation. The analysis of travel patterns should ultimately inform planners in making decisions ranging from strategic and tactical to operational planning. This can be further facilitated by fusing long- and short-term predictions of travel patterns. While most of the studied included in this review have alluded to potential relevant applications, none of them has made an explicit link to service planning tasks, with the exception of Yap et al. (2019) and its relevance for service coordination.

There is thus clearly a missing link between the development and application of demand, urban and network clustering analysis on one hand and service design and decision support on the other hand, which is key for the value proposition of the literature reviewed here. Future research should bridge this gap by making explicit how the techniques and insights developed based on smart card data analytics can translate into service improvements. The latter can range from ex-ante policy assessment (see Kholodov et al., 2021 for an example concerned with price elasticities) and network and capacity allocation adjustments (see Kolkowski et al., 2023 for an example concerned with urban activity segregation impacts of a network development) to the design of flexible services (see Qiu et al., 2019 for an original contribution in this direction) on one hand to real-time management such as information provision and control measures on the other hand. In particular real-time control strategies will greatly benefit from advancements in short-term passenger flow predictions. Another promising direction is the development of fare products and fare schemes (see Halvorsen et al., 2020 for a relevant framework) based on the market segmentation performed using smart card data. For example, changes in travel patterns induced by the COVID pandemic crisis and work-from-home habits call for the development of new subscription models which can be devised to cater for changes in spatiotemporal user clusters.

Smart card data which originated and has been designed as a means of facilitating automated fare collection has clearly emerged as an invaluable source for analysing human mobility patterns. Fusing smart card data with other sources of geo-location data will potentially allow analysts to consider additional aspects of human mobility beyond those performed by means of public transport. Moreover, for the research analysing mobility patterns using smart card data to become an integral part of service design and decision making and offer insights that go beyond describing current travel patterns and related user segments, future research must combine methodological advancements in data science, machine learning and clustering techniques with knowledge in travel behaviour, demand forecasting and transport planning.

Disclosure statement

No potential conflict of interest was reported by the author(s).