Spatiotemporal variation in travel regularity through transit user profiling

At coarse spatial and temporal scales, urban transportation systems might be assumed to be an amalgam of highly routinised traveller behaviours. During each day, barring major events or disruptions, travel demand rises and falls in broadly predictable ways, predominantly aligning with patterns of work-related travel at the beginning and end of the working day and from suburbs to city centres and core workplaces and back. Yet substantial explorations of travel behaviour have shown that the nature of such regularity is considerably more complex, varying within and between individuals on a day-to-day basis (Hanson and Huff 1986), influenced by socio-economic status and land use activity types (Zhong et al. 2015), trip chaining (Primerano et al. 2008), changes in weather (Saneinejad et al. 2012), and simple ad hoc requirements such as managing congestion through physical means. Rather than assuming the widespread persistence of routine behaviours, transportation system dynamics can be better understood by capturing how regularity of persists at the individual scale. Through various measures of individual regularity with respect to the occurrence of trip-making, it is possible to explore how these dynamics can vary widely over space and time.

Regularity in travel behaviour has been an active research area ever since the first surveys of household transportation usage were carried out over half a century ago. In the 1970s and 1980s, diurnal activities which generated trips in households within the framework of time budgeting led to many diary-based surveys which demonstrated the existence of day-to-day variability in travel activity, and thus highlighted limitations in trip generation models built using peak hour or entire day survey approaches (Hanson and Huff 1986; Huff and Hanson 1986; Hanson and Huff 1988; Pas 1987; Pas and Koppelman 1986; Jones and Clarke 1988). Variability in departure times (Mahmassani and Chang 1985), route choice (Mahmassani and Stephan 1988), mode choice (Mannering et al. 1994), the regularity activity engagement (Minnen et al. 2015), and temporal trip rates (Pas and Sundar 1995) were investigated. More recently, mobility data through GPS traces has enabled the automation and extension of longitudinal travel behaviour analysis (Pendyala 1999; Axhausen et al. 2002). Through these data, measures of social travel (Schlich and Axhausen 2003; Schlich et al. 2004), the extraction of activities and land use related to trip-making (Huang et al. 2010; Zhong et al. 2015) and the definition of individual activity spaces (Schönfelder and Axhausen 2003; Schönfelder et al. 2006) have been explored.

While conventional approaches yield useful insights into the nature of regular travel behaviour, the typical scale of data collection means they inevitably fail to capture wider variation in behaviours induced by the social, spatial and environmental context (Van Acker et al. 2010). At present, there is little understanding of large-scale spatiotemporal variation in regularity behaviour. Newly available, large-scale mobility datasets do, however, provide significant opportunities for widening insights into travel behaviour. These datasets capture the mobility of large proportions of the travelling population at fine spatial and temporal scales, usually focussing on the homogeneity and heterogeneity of traveller behaviours at many temporal scales but all constructed from the very fine scale data that is streamed in real time captured from automated devices. In recent years, travel behaviour research has been strengthened through analysis of large volumes of smart card transaction data (Morency et al. 2006; Pelletier et al. 2011; Long and Thill 2015; Zhong et al. 2016), GPS route traces (Manley et al. 2015; Wu et al. 2014), and mobile phone movement data (Gonzalez et al. 2008; Schneider et al. 2013; Ahas et al. 2015). Yet less well researched are the opportunities these data afford in analysing longitudinal behaviours. Through collection of mobility trends at the level of the individual traveller, new potential for profiling travellers and observing changes in behaviour over time have emerged. The analysis of traveller profiles from such data is the main goal of this paper.

Although data volumes in transportation analysis tend to be large due to the fact that the size of potential movements between many locations has always been substantial long before the age of real-time streamed data, it has not been possible to capture the kind of diversity in trip-making that is associated with many spatial and temporal scales until quite recently. Transportation data such as that generated by individuals activating sensors which capture their locations and movements enables behaviour at all scales to be examined and only now is it possible to define and visualise the enormous heterogeneity in traveller behaviour from such datasets. In short, this kind of ‘big’ data forces us to recognise that regularities and irregularities which are related to extreme events as well as variations in individual behaviour must be separated out from more general noise in such data if we are to produce a better understanding of the reasons why and when people travel.

Regularity is thus only interesting if we are able to define this as a series of baselines against which to compare more ad hoc travel such as that pertaining to disruptions, infrequent and extreme events, individual variations in generic travel behaviour and variations across different modes, temporal and spatial locations. Such an understanding is essential for policy making which must build on the deepest understanding of travel dynamics that we can determine. We also need to determine the regularities associated with individual travellers and their aggregation into groups of like-behaviours. In short we need to explain how individual regularities add up to collective regularities and the way such regularities become irregular as we aggregate and disaggregate over space and time. We need to achieve this kind of understanding so we can target policies for improving the operation of the system and the travel experience for passengers at the right spatial and temporal scales. As we will see, the data from the smart cards used for fare collection on public transport in London (called the Oyster card) is able to provide us individual-level long-term views of traveller behaviours. This provides a potential view on travel regularity at spatial and temporal scales that have not been previously possible through conventional datasets. These insights can yield benefits for transport planners, where differentiation between regular and ad hoc travel can contribute towards shaping how services are planned and operations amended, responsive to different types of travel behaviour.

As we have individual data on where a person begins and ends a trip, we can search for regularities across any spatial and temporal scale with respect to the volume of travellers at different points in space and time–stations and times of entry and exit. Such as in (Gong et al. 2012), peak hours and differences between weekday and weekend are analysed using smartcard data from Shenzhen. Zhong et al. (2015b) further extend such analysis from three levels-trips, stations and urban system, demonstrated using data from Singapore. More examples can be found from (Pelletier et al. 2011; Yue et al. 2014). There are literally thousands of possible kinds of regularity to be explained when we have individual data at the most micro level, but in this paper, we will restrict ourselves to regularities that pertain to particular windows of time and to locations, not to aggregations of different windows or to sets of locations which imply actual trips between origins and destinations. In fact, we will restrict our analysis in this paper to regularities associated with entries and exits from the transit systems but not regularities associated with actual trips where we need to match exactly entries with exits for individual travellers, nor regularities that pertain to trips by a single individual during the day such as commuting to and from work.

To determine regularities, we use a very straightforward technique of clustering individual behaviours in terms of when they take place over days and in particular locations using criteria that is plausible with respect to evaluating the degree of similarity between events. The method we use is called DBSCAN which was first developed by Ester et al. (1996) and is a simple method of relating events—where in this case a traveller registers at an origin entry or destination exit—which are close to one another in terms of their temporal distance from one another. The method also assumes that there is a minimum size to the cluster and in this way, a simple algorithm is able to extract unique clusters from any dataset whose events are ordered in linear time—such as the sequence of tap-in or tap-outs on a transit system. The DBSCAN method has been used for clustering travel data in Brisbane (Kieu et al. 2014) and in Beijing (Ma et al. 2013) which provide good comparisons with the analysis of profiles here. This paper extends these earlier applications by measuring how regularity unfolds across space and time. For capturing clusters in our transit data, we will detail how the method works in detail when we discuss the analysis below.

Thus the purpose of our preliminary analysis is to analyse large-scale variation in regular and irregular travel behaviour. Building upwards on individual-level smart card travel transaction data, it is possible to derive a system-wide spatiotemporal understanding of regularity in travel behaviours. In the first instance, the paper will describe the dataset and context within which the study is conducted. Following this, “capturing traveller regularity” section provides a description of the computational method used to capture individual-level regularity, outlining three different definitions of travel regularity. Using these definitions, the paper undertakes two stages of analysis. “Regularity at different scales” section addresses the nature of regularity according to each definition, examining its persistence and nature across all travellers over space and time. “Spatiotemporal variation in regularity and irregularity” section goes deeper into the existence of regularity relative to non-regular behaviours, providing insight into the causes of regularity, showing the influences of demographics, land use, work place patterns, and major events. The paper will conclude with a discussion of how this view of travel behaviour can aid in planning and modelling the transportation system, and how these methods might be refined and extended in future research.

This study is based on an archive of public transportation smart card transactions data recorded for public transport in Greater London (UK) between June 16th 2012 and September 9th 2012. Smart cards—known in London as Oyster cards—allow travellers to enter and exit the transportation system having paid their fare prior to use of the system, these monies being recorded on the smart card. The Oyster card system is used by around 85% of all travellers on the London Underground and bus network although this percentage varies during the day as it is still possible to purchase traditional paper tickets for use on National Rail and London Transport outside the London region (Transport for London 2013).

As travellers enter and exit the system, transactions are recorded that provide details about the type of trip that has been undertaken. Each transaction record includes references to type of card, travel mode, date and time of travel, trip origin or destination (in the cases of rail and underground travel), and the service choice (where trains or buses are taken). These data are recorded for each traveller, denoted by a unique, but anonymised, identifier. The Oyster card is little different to the other smart card data sets that introduced in (Agard et al. 2006), and more details can be refer to (Gordon et al. 2013).

Given the personal nature of the data, no additional metadata is provided in relation to the traveller holding the card. This limits the potential depth of analysis, for instance, travel behaviour relative to home and work locations has to be derived through inferring methods (Zhong et al. 2014); demographic attributes to a user has been enriched through integration of the other data sets (Long and Thill 2015), however, its reliability needs further investigated given the relative sparsity of the network. Unlike these approaches, this paper deepens the analysis focusing on making use of the best features of the data itself (long-term, large volume and good coverage) through a systematic framework.

The study uses transaction data recorded across 49 weekdays over the approximate 3-month period including the 2 weeks of the Olympics Games in London which has been excluded from this study. The dataset contains 639,979,438 transaction records, averaging at around 13 million transactions per day, for 11,535,090 individual travellers. Transactions are classified into ‘tap-in’ events at the start of rail journeys, and located at a specific rail station, ‘tap-out’ events, which mark the end of a journey (and can be linked to ‘tap-in’ events), and bus journey events, which are provided with the bus service identity but are not spatially located. Table 1 provides a breakdown of the occurrence of each frequency type.

The purpose of our research is to identify and analyse the regular behaviour of each distinct traveller. As we noted above, regularity in travel behaviour can take many forms, dependent on simply travel time, travel mode, or a specific location, so it is important that the methods used in identifying regularity remain flexible to these variations. The method should impose as few assumptions on how regularity is manifested as possible. In order to avoid placing specific constraints on the nature of regularity, the simple temporal clustering approach noted above is adopted. Clustering allows us to identify high density events in temporal activity, indicative of regular behaviour. Various approaches can be adopted, but the method most closely aligned with the requirements stated above is DBSCAN (Density-based spatial clustering of applications with noise) (Ester et al. 1996). DBSCAN places few constraints on how a cluster must appear by building clusters based solely on the spatial density of data points, in contrast to other clustering methods that require a predefined specification of cluster size (e.g. k-means clustering) or assume hierarchical linkage between clusters.

The DBSCAN method requires the specification of two parameters, both of which are related to the required density of data points that constitute a cluster. The first is a definition of the minimum number of points (minPt) that a point must be near to in order for the point to be deemed to be within a cluster, and the second is a specification of the temporal distance (dist) threshold within which two points would be classified as near to each other. Although different measures have been tested, for brevity this paper will not explore the impact of variations in these parameters on cluster sizes and definitions of regularity, we consider these values that we have chosen to be optimal. Through these tests, a combination of dist = 30 min and minPt = 5 produced clusters that aligned well with observations within the data of regularity and repeated journeys, producing intuitively acceptable and plausible clusters of events. While these parameter settings may be applicable elsewhere, it should be noted that these were set specifically for the range and size of the study dataset based on traveller behaviour in Greater London. As we show elsewhere in a comparison of Beijing, London and Singapore (Zhong et al. 2016), there are differences in diurnal variability of travel volumes in these different cities and it is likely that these would be reflected in the parameters which produce optimal regularities. We have not tested for this variability on these different datasets as yet.

The specification of the clustering algorithm is only one element, however. Consideration must also be given to the type of travel behaviour being clustered. Within the Oyster card data, a differentiation is made between trip starts (or ‘tap-ins’) on the Underground and rail system, trip ends (or ‘tap-outs’) on the Underground or rail, and bus journeys which only register the event of being on a bus. Each record has a time associated with it, and a station or bus service at which the travel record was generated. In order to again retain flexibility in our analysis of regularity, it was decided that clustering would be applied in three ways—firstly, to all transactions, regardless of type; second, by transaction type, regardless of location; and third, by specific transaction at a specific location (station or service). This differentiation provides variation in the specificity of the regularity definition, maintaining flexibility in the definition, but furthermore enables insight into how far regularity persists despite increasing specificity. As we implied earlier, we consider these to be the most fruitful elements of regularity in this preliminary study.

Figure 1 explains how clusters are formed using DBSCAN through use of each subset of the data. As can be seen, clusters are formed according to a threshold of point density, which varies according to the data incorporated. By including all transactions, a larger cluster is formed, spanning a wide timespan. Greater specificity in the subset of the data—either by mode or service—naturally leads to a reduced cluster size. In all instances the DBSCAN parameters govern the formation of clusters, excluding any points outside of the dist threshold, and clusters of points not exceeding the minPts parameter (most clearly seen through absence of cluster of Station B points in the Service-Temporal example). Readers are referred to the paper by Ester et al. (1996).

Regularity clusters will be generated from the trip records of each individual traveller. By extracting regularity at the individual scale, it will be possible to aggregate across spatial and temporal definitions, providing a detailed view of regularity across the transportation network. It will furthermore enable the extraction of linked regularity clusters, highlighting where interdependencies between different types of travel behaviour exist. Details of how regular interactions are determined will be provided below.

The first stage of our analysis examines the nature of regularity across each or our three definitions Having derived regularity clusters for each traveller for each definition, this analysis will describe their appearance, structure, and persistence over time. Following the description of each form of regularity cluster, we will then describe interactions between cluster types.

While trends in regularity provide some insight into travel behaviour, regular behaviours can be better understood relative to all travel. The intention of the second stage of this analysis is to therefore provide insights into the persistence of regular travel relative to irregular, ad hoc travel. These insights should provide a view of the type of regular behaviour unfolding at different places and times across the transport system. There are three major concerns underlying this phase of the analysis, first to establish the nature and degree of regularity across different transport modes, second to observe how regularity varies across space and location, and third to explore how regularity varies across different time periods. In each case, the volume of regular trips, defined using the above approaches, are measured against all comparable trips.

Our analysis so far has been largely descriptive but we believe that such detailed scrutiny of big data is an essential prerequisite to much more relevant and deeper understanding of travel behaviours than the current state of the art and practice reflects. We have known for a very long time that the pre-morning and morning peak are very different from the evening peak but it is very clear from our analysis here that within these broad trends and patterns, there are considerable departures from the kind of regularity that are assumed in existing models of transportation behaviour. None of this is very surprising given what we know about the increasing complexity of travel behaviour in cities with workers having ever greater flexibility over their working hours and places of work, as well as the growth of tourism in world cities which accounts for an increasing proportion of day time trips. Moreover, new patterns of significant transport to health care facilities and to schools distort the traditional morning and evening peaks, and although these patterns might show their own regularity combined with work and tourism, the whole picture can become confused. In short, different regimes of regularity can combine to generate what might appear to be quite irregular sequences of travel events. This simply points to the notion that we need other datasets associated with geodemographics and related socio-economic characteristics to give added value to this new kind of real-time streamed transit data so we can ground the patterns we have found here in less ambiguous modes of explanation. This challenge sets up a research agenda for the coming years.

We also need to move beyond examining simple clusters of travel events, toward events that are formally linked in space and time. For example, if an individual produces a cluster of events within a short time over a long period which might be interpreted as the origin events defining a journey to work, then we need to match this to destination events which might be part of a different cluster which define the concluding point of the journey to work. This kind of repetition is much harder to recognise especially if other trips are mixed with this such as visits to retail facilities and so on. Going further and matching clusters in the morning peak with the evening or at any times of the day is even more difficult to interpret as a reverse journey from work to home but it is interactions such as these which are essential to a comprehensive understanding of travel behaviour. These are all goals for the future and there are considerable possibilities of taking this analysis further using datasets of the kind we have worked with in this study. Only when we generate a much more complete understanding of regularity across all scales which involve aggregation of space/locations, times, and numbers of individuals will we be in a position to build multilevel models that begin to address the real complexity of travel behaviour. This is the message that Kitamura et al. (1997a, b) have been arguing for a number of years.

In progressing this agenda, we also need to be wary of questions involving confidentiality in individual based datasets such as the ones we have access to. Transport for London who collect and archive the data that they have provided us with have strong safeguards in place and it is impossible to identify any particular individual. But if we had a detailed classification of every kind of transportation behaviour there is, then this would be easier. We have assumed so far that every set of travel events associated with an individual in the data is unique although we know this is not the case especially once we begin to aggregate data so that the truly random occurrences associated with time when travellers tap-in and tap-out are smoothed. A bigger issue is the fact that some of the data is scrambled because there is no guarantee that each individual identifier is unique in that often individuals have more than one Oyster card and to all intents and purposes, these imply different travel behaviour. Combined with considerable loss of data due to free cards which do not require a tap-in or tap-out, as well as the conflation of season ticket holders with daily travellers, this adds to what essentially we can only interpret as noise in the data. Last but not least, we need more detailed analysis of different types of travel behaviour which are specified a priori and then used to determine whether such types exist in the data. Currently our approach is entirely deductive as is much analysis of big data – which is based on the search for pattern—rather than testing whether or not different hypotheses about how people travel can be discerned in the data. We need to extend our analysis in a more deductive manner so that we might redress the balance and introduce more explicit hypothesis testing in explaining travel behaviour.

Research around regularity should extend more widely beyond the methodology introduced in this paper. This work has focused on Oyster Card transactions alone, and therefore limited in its description of all transportation behaviours. It is likely that Oyster Card users, by virtue of the lengthier process to buying paper tickets on the London transportation system, are more regular than the remainder. One furthermore is able to gain little view into the regularity of private car drivers, cyclists, and multi-mode users. Advances in relation to these issues are becoming viable, given the movement towards universal payment through contactless payment cards (replacing both paper tickets and smart cards), and integration of transport modes through MaaS (Movement as a Service) platforms. In the shorter term, however, as has been shown elsewhere it is possible to apply these methods to other forms of spatiotemporal movement data, such as mobile phone traces. Other datasets, should they become available, may extend and support patterns derived from smart card transactions alone. While this paper has focused on the descriptive potential of regularity analysis, alternative methodologies should be explored. Advances potentially lie around alterative configurations of DBSCAN or investigations using other clustering methods for capturing individual-level regularity.