Computers, Environment and Urban Systems A spatiotemporal and graph-based analysis of dockless bike sharing patterns to understand urban ﬂ ows over the last mile

The recent emergence of dockless bike sharing systems has resulted in new patterns of urban transport. Users can begin and end trips from their origin and destination locations rather than docking stations. Analysis of changes in the spatiotemporal availability of such bikes has the ability to provide insights into urban dynamics at a ﬁ ner granularity than is possible through analysis of travel card or dock-based bike scheme data. This study analyses dockless bike sharing in Nanchang, China over a period when a new metro line came into operation. It uses spatial statistics and graph-based approaches to quantify changes in travel behaviours and generates previously unobtainable insights about urban ﬂ ow structures. Geostatistical analyses support understanding of large-scale changes in spatiotemporal travel behaviours and graph-based approaches allow changes in local travel ﬂ ows between individual locations to be quanti ﬁ ed and characterized. The results show how the new metro service boosted nearby bike demand, but with considerable spatial variation, and changed the spatiotemporal patterns of bike travel behaviour. The analysis also quanti ﬁ es the evolution of travel ﬂ ow structures, indicating the resilience of dockless bike schemes and their ability to adapt to changes in travel behaviours. More widely, this study demonstrates how an enhanced understanding of urban dynamics over the “ last-mile ” is supported by the analyses of dockless bike data. These allow changes in local spatiotemporal interdependencies between di ﬀ erent transport systems to be evaluated, and support spatially detailed urban and transport planning. A number of areas of further work are identi ﬁ ed to better to understand interdependencies between di ﬀ erent transit system components.


Introduction
Cities are complex systems, composed of people, places, flows, and activities (Batty, 2013). Quantifying their dynamics, system interdependencies and spatial structures can characterise urban morphology and metabolism. People as physical carriers, drive the flows of materials, money and information within urban spaces, and influence economic growth, social equity (Batty, 2013). Understanding the nature of these flows provides perspectives and insights into how socioeconomic and environmental problems such as urban development, transportation efficiency and air quality are being addressed (Borrego et al., 2006;Desouza & Flanery, 2013;Fishman, Washington, & Haworth, 2014;Gong, Lin, and Duan, 2017).
Travel data can be used as proxies for urban flows because they describe people's movement. Traditionally, such data were obtained from household travel surveys, with high cost and time overheads.

Background
Bike sharing schemes have become increasingly popular in recent years, reflecting their environment friendly, low cost and convenient nature. They are understudied, with research focused on cycling behaviours associated with dock-based bike sharing schemes. Vogel, Greiser, and Mattfeld (2011) examined geographical clusters of docking stations using spatiotemporal usage. Others have identified bike fleet rebalancing strategies for different types of stations and quantified links between bike demand and land use (Daddio & Mcdonald, 2012;Jiménez, Nogal, Caulfield, & Pilla, 2016;Kaltenbrunner, Meza, Grivolla, Codina, & Banchs, 2010;O'Brien, Cheshire, & Batty, 2014).
However, bike sharing schemes can play important role in examining the "first/last mile" problem. This is the distance between home/workplace and public transport that is too far to walk (Fishman, 2016;Saberi, Ghamami, Gu, Shojaei, & Fishman, 2018;Shaheen, Guzman, & Zhang, 2010), and bike schemes provide access to other forms of public transport and mass transit (train, metro, bus etc): they act as the "capillaries" for the mass transit aorta. The advent of dockless bike schemes opens up the opportunity to examine the last mile in detail.
To understand the last mile using bike data, the provision of other transport systems needs to be considered as well. Many studies have examined how cycling and metro trips are combined (Lin, Wang, & Feng, 2017;Martens, 2007), how this varies for different socio-economic groups (Zhao & Li, 2017), are affected by pricing  and has sought to quantify the interdependencies between bike sharing schemes and metro systems (Ding, Zhang, & Xie, 2019;El-Assi, Mahmoud, & Habib, 2017;Ma, Liu, & Erdoğan, 2015). Most of these studies have found a positive correlation between metro stations and bike sharing trips, but some have questioned this (e.g. Tran, Ovtracht, & D'arcier, 2015). Overall, these studies have focused on bike sharing preference, bike trip spatial clustering around other transportation hubs (metro, tramway and railway stations), and have ignored flows and structures in the last-mile. Examining the relationships between bike and transportation flows and structures can lead to deeper understanding of urban dynamics. Saberi et al. (2018) analysed spatiotemporal statistics and network (graph structure) properties of bike sharing trips to examine the impact of metro strikes, identifying increase in bike use (numbers and trip distances). Chen et al. (2016) constructed a framework to predict the short-term over-demand periods for sharing bike station clusters considering of metro delay. Both studies used data from dock-based bike sharing schemes. These have a number of important shortcomings including service coverage (e.g. bikes may not available in the suburbs), docks may largely distanct from the actual origin/destination (OD) and there can be dock capacity/availability issues. Thus the inference derived from analyses of these data are limited.
The development of the IoT (Internet of Things) has the potential to revolutionise many aspects of our lives which are increasingly connected and sensed, generating large volumes of data with location and temporal attributes. In bike sharing schemes, dockless bikes emerged around 2015 and rapidly became a success in number of countries including China (iiMedia_Research, 2017;Xu et al., 2019). Unlike traditional bike sharing schemes where bikes can only be borrowed and returned at docking stations, dockless schemes enable users to locate and borrow bicycles via a smartphone app, returning them to any suitable public location. Dockless schemes allow convenience and flexibility for users. The smart lock system and GPS unit on the bikes not only facilitate scheme operation and bike management, but also create a large quantity of spatiotemporal individual level data. An advantage over data from traditional dock-based bike sharing schemes, is that the flows captured by the data are more detailed (i.e. with higher spatial granularity) and better capture people's actual activities, travel demands and behaviours.
Dockless bike sharing studies are few and have focused on bike fleet management (Pal & Zhang, 2017), planning of related infrastructures (Bao, He, Ruan, Li, & Zheng, 2017), and bike distribution patterns (Liu, Shen, & Zhu, 2018). Liu et al. (2018) proposed combining a factor analysis and convolutional neural networks for inferring dockless sharing bike distribution in new cities. This work found these spatial distributions are correlated with nearby urban context, and change over time. Other studies (Ai et al., 2018;Xu et al., 2019;Yang, Heppenstall, Turner, & Comber, 2018) have shown that bike distribution is significantly time dependent, especially around metro stations, suggesting that dockless bike flows and activities are highly dynamic. Zhang, Lin, and Mi (2019) developed a framework for planning geo-fences to constrain dockless bike parking. Clustering analysis and location-allocation models were applied to assess the implications of spatial planning of geo-fences in different scenarios. While this framework incorporated spatial detail, it lacked explanatory social and economic information, for example related to the cost of geo-fences and punishment/reward of parking bikes outside/inside geo-fences. Zhou, Ni, and Zhang (2018) used questionnaires to examine attitudes over the effects of dockless bikes on metro commuting. Their results showed a significant positive influence, especially for non-motorised vehicle owners and metro stations outside city centres and highlighted the positive role of such schemes in mass transit systems.
Analyses of data from dockless bike schemes have the capacity to provide high resolution insights into people's non-motorised mobility patterns and behaviours and to reveal their relationship with other urban structures and processes, for example flow networks in other mass transit system (e.g. metro), urban infrastructure development (e.g. new train stations, new bridge) and related urban updates. Critically such data allow such relationships to be examined dynamically over the last mile.

Study area
Nanchang in southeast China is the capital city of Jiangxi province with a population of 2.15 million. The city has a typical humid subtropical climate, characterized by hot and humid summers and mild winters. It has two metro lines as of September 2017. A new line of 17 stations, Metro Line 2, opened and started running on August 18, 2017. Fig. 1 shows the transit map of Nanchang with the Gan river running through the city and the two metro lines. Other public transit systems include bus and dockless bike sharing. This study analysed data of around 80,000 dockless bikes from the Nanchang urban area around the time of the opening of Metro Line 2, specifically to compare bike usage "before" and "after" the opening of the new metro line.

Data description
A program was set up to collect dockless bike availability data via the bike scheme API (application programming interface) for the month of August 2017. Queries to the API can return bike availability for any specified location (point) returning information on bike identifiers and their coordinates, with an in-built limit of the nearest 30 available bikes. The program iterated through the whole urban area collecting data on bike availability on a raster grid of 0.0015 degrees (length of sides equals to approximately 150 m). Most available bikes locations across Nanchang were captured approximately every four minutes due to the large urban area and the API query limits. There were some gaps in coverage due to bike GPS and communication unit signal receiving problems (e.g. GPS does not function well in tunnels). The data include bike ID (identifier), coordinates and the timestamp information. Fig. 2 (a) shows a snapshot of dockless bike in Nanchang, with each point representing an available dockless bicycle.
Extreme weather conditions have been found to have significant negative effects on cycling activity (El-Assi et al., 2017;Zhou et al., 2017). To ensure consistency across both time periods, only data from rain-free days were analysed, with weather data from wunderground. com (one meteorological station, every 3 h), and although other conditions (e.g. air condition, temperature) may have minor impacts on bike usage, these were relatively consistent over the study periods. Data for 5 weekdays before and 5 weekdays after the opening of new metro line were selected for analysis, with all data collected in the same month and under the same general weather conditions in rain-free days. Table 1 shows an example of the data collected and Table 2 shows   the dates of the data used in the analysis. Open Street Map data were also used to provide basic background mapping and road network data.

Detecting bike trips
The data provide snapshots of available dockless bike locations, bike identifier and timestamps. From the stacks of snapshots, it is possible to identify the changes in location of an individual bike and thus to derive dockless bike trip information. For this study, the threshold trip value was 100 m: if the change in location of an available bike between two timestamps exceeded this threshold, then the two records were linked as a trip, with the earlier one providing the origin coordinates, and the latter one the destination. It should be noted that this method is unable to identify round trips where the origin is close to destination. Fig. 2 shows an example trip. The bike changes its location from t1 to t5, which are combined to form a bike trip. It stays in the same place from t5 to t9 and is considered as available in this period. Similarly, a second trip can be identified from t9 to t10. The dockless bike sharing scheme also allows bikes to be scheduled (booked) in advance of 15 min, but this still allows trip origin and destination to be identified, although with a small impact on trip duration.

Constructing dockless bike mobility graph (network) structure
Converting the bike data into a graph structure to represent flows allows the changes arising from the introduction of a new metro line to be examined. Typically, any system composed of interconnected individuals can be viewed as a graph (i.e. network) with individual components are represented by nodes (i.e. vertices) and their interconnections by arcs or links (i.e. edges). Examples include online friendship networks, scientific collaboration networks and global airline routes. In urban and transportation studies, bike sharing docking stations (Austwick, O'Brien, Strano, & Viana, 2013), bus stops and metro stations are typically viewed as nodes. If there is at least one trip between a pair of nodes (stations), then a link is generated between them. By representing the relationship between transportation nodes (e.g. bike/bus/metro stations) as travel flow links, mobility graph structures can be constructed. For example, Saberi et al. (2018) studied the dynamics of bike sharing networks in London, casting docking stations as graph nodes, and flows between stations as links. However, dockless bike sharing schemes provide greater spatial detail about the patterns of bike movements and therefore their riders than docking station data. Because bikes and bike trips are not aggregated over docking stations, constructing graphs of dockless trips requires different considerations. Fig. 3 (a) shows an example of bike trip origins in an urban area. Instead of being located at fixed docking stations, they are more loosely distributed along roads, and sometimes clustered at certain road segments. Fig. 2 (a) also confirms that majority of the bikes are located close to roads, with few of them distant to the road network in the north and southeast. These are locations outside the Nanchang urban area, and distant from metro system. This study uses road segments to aggregate bike locations and to cast them as vertices in a graph structure. The graph links represent trips originating from one road segment to another. Choosing the right spatial scale of aggregation is important because this can impose a source of bias in spatial analysis (Openshow, 1979). Some research (Calabrese et al., 2011;Louail et al., 2014) using cell phone data to detect commuting flows in cities choose grid cells to aggregate flow origin and destinations with size varying from 500 m to 2 km. These scales may be useful for capturing work-home commuting flows and inferring coarse scale urban structure, but here the focus was on identifying more spatially detailed structures in small areas requiring a finer spatial scale. Fig. 3 (b) reveals that trip distances are commonly around 400 m and that 28% of trip O-D distance are < 500 m, with 1 km and 2 km distances corresponding to 60% and 85% or trips respectively. This indicates that aggregation scales between 500 m to 2 km may result in important information loss. Here, road segments of 200 m were used to aggregate dockless bike flows, since this distance captures > 96% of trips. Fig. 4 illustrates the process of casting bike trips to a graph using road networks segments as nodes. First, the Nanchang road network was split into segments based on road joins. Second, if a split segment was longer than 200 m, then was divided into smaller segments of equal distance, all < 200 m. Then, bike trip origin and destination were assigned to the nearest road segment using ArcGIS 10.4. The result is that each trip has information about its origin and destination road segment, with a trip regarded as a flow from one node to another.
Dockless bike travel flows and the graph structures in the "before" period (5 days) are shown in Fig. 5, using a spatial layout and a Fruchterman Reingold layout. Each node represents a short road segment with the edge between any two nodes indicating travel flow between them, and the network was trimmed using a threshold of at least 10 trips between two nodes, to improve the visualisation. Fig. 5 characterizes nodes degree (number of connections) using both shading and node size. Node degree in an urban mobility network represents the connectivity or accessibility to destinations or activities across the network. In this case study, bike travel is split by the river, with few links crossing it (Fig. 5 a). Most of the trips are local, with the majority of the links connecting nearby areas (Fig. 5 a), and the graph structure has a clear multi-core (multi cluster) spatial pattern. This is different to graph structures derived from other transportation modes such as metro and bus travel (e.g. Zhong et al., 2014) which tend to link different parts of city over longer distances. The Fruchterman Reingold layout (Fig. 5 b) seeks to reduce the overlaps between nodes and maintain spatial topology. For example, two clusters of nodes presented in

. Temporal pattern
Temporal analysis of bike usage can reveal the dynamics in and characteristics of dockless bike sharing schemes. Fig. 6 (a) shows the temporal travel patterns of dockless bike trips in Nanchang (whole study area) with the highest daily temperature. Trip amount is the count of trips starting at different times in hourly intervals. Over the two periods, bike usage shows a similar overall temporal pattern with some difference in trip numbers. There are two significant rush hours in each day, one from 07:00 lasting for two hours and another around 18:00, with hourly trip number reaching approximately 19,000. There is also a lunch time peak at 12:00. The trip amounts overlap with the exception of Monday at 12:00, when it is slightly lower, which is potentially due to the hotter weather (38°C) than any other days ( Fig. 6  (a)). The total trip amounts for the two periods is 838,464 and 892,764 respectively, an increase of 6.5%.
For the purposes of illustration Fig. 6 (b) defines service area    catchments of 2 km around new metro stations. Trip amounts in these catchmentss increased by 28.0% from 96,239 to 123,182. The highest rise was found in early morning with increases of 38.5%, 38.0% and 31.3% in the hour intervals from 06:00-09:00, respectively and another increase around 20:00 by 34.7%. The observations indicate that the largest bike trip amount increase occurs in the morning rush hour period, suggesting the popularity of trips using the new metro with dockless bike as a preferred travel mode during morning commuting.

Spatial pattern
Dockless bike sharing scheme offers flexibility, efficiency, and low cost making it an attractive way to fill the gaps in the public transit system (such as bus and metro). Bike sharing travel has different spatial properties from other public transit modes, due to its non-motorised nature. For example, dockless bikes are not suitable for long distance travel, but they can be left immediately adjacent to the destination. Understanding the unique spatial pattern of dockless bike usage is important for scheme management (bike re-balancing) and the interdependence between metro and dockless bike system provides crucial information about local mobility. A spatial analysis (Pebesma, 2018) was conducted to examine the characteristics and dynamics of dockless bike sharing scheme in the study area. Fig. 7 (a) shows the dockless bike trip kernel density across the city in the five weekdays of the before period based on trip origins. A KDE bandwidth of 118.2 m was determined as follows:: where SD is the standard distance of the points spatial distribution, D m is the median distance, and n is the number of points. The KDE has a 100 m resolution. The city centre from where most to trips originate, straddles both sides of the river,. It has the highest kernel density value (21,471 trips per km 2 ). Several hubs can be also observed in the south away from city centre and the metro service. Fig. 7 (b) shows the changes of kernel density estimation (KDE) between the two periods. The highlighted area in Fig. 7 (b) shows the new metro service catchments west of the river. It shows a clustering pattern of increased density. By contrast, in other part of the city, areas of increase bike use are adjacent to or nearby to areas of decrease, suggesting an overall balanced and dynamic pattern. The observed spatial proximity between increase and decrease is due to the nature of dockless bike sharing scheme: people are able to pick up and park bikes at places of their choice when necessary subject to availability. The flexibility of travel in this way is demonstrated by heterogenous use patterns, especially at fine spatial scales.
To better understand the impact of new metro service on dockless bike sharing trips, a further statistical analysis was conducted to examine the probability density of dockless bike origin/destination spatial distribution. Fig. 8 (a, b) illustrates the probability and cumulative probability of start and end distances from the newly opened metro stations.
By comparing Fig. 8 (a) and (b), a new peak at around 120 m in the "after" period is observed, which suggests that more bike trips originate from or end at areas very close (around 120 m) to metro stations. This distance can also be understood as how far metro travellers are typically prepared to go (walk) to get or leave bikes around their metro usage. The first peak (Fig. 8 a) decreases and reaches a floor around 220 m, indicating weakened spatial clustering of bike trips origins and destinations. This distance can be interpreted as the affordable (walking distance) limit between the metro station and the bike, and provides information to support the planning of related bike parking facilities, suggesting in this case an upper limit for the provision of bike parking areas around metro stations. The highest peak in the "before" period is located at around 800 m, but moves to 600 m in the "after" period ( Fig. 8 a), suggesting the impact of the new metro stations on bike usage. The cumulative probability (Fig. 8 b) indicates however that this impact is heterogeneous, with higher cumulative probability in areas at shorter distances to new metro stations in the "after" period. Figs. 7 and 8 indicate the extent to which the areas closer to new metro stations experienced greater changes in daily dockless bike trips, in terms of distances people walked to get a bike and volume of trips.
Metro services not only change where people pick or park their bikes, but also influence local trip distance patterns. Building on the results in Fig. 8 (a), buffers of 250 m and 2,000 m around new metro stations were used to select bike trips with origin or destination within them. Trips that start or end in the 250 m buffer can be understood as "last mile" trips, and trips in the 2,000 m buffer capture flows within the metro service catchment area, indicating changes over larger areas. The travel distance patterns were examined their distributions compared (Fig. 9). Fig. 9 (a, c) indicates that bike trip counts increased significantly in both buffer zones and that the new metro service stimulated more trips regardless of travel distance. Fig. 9 (b, d) describes the changes in probability distributions and indicates that trips with a travel distance of < 1,000 m are more likely to be made with the introduction of the new metro service and long distance trips (> 1,000 m) less likely. t-tests were used to confirm the difference in dockless bike trip distances in the two periods (Table 3)  for "last mile" trips, and trips in the larger service areas around new metros (radius of 2,000 m) significantly decreased from 1,317 m to 1,150 m. Overall, the results suggest that new metro stations encouraged increases in cycling activity but with reduced trip distances, indicating improvements in local mobility and access to transportation services.

Graph-based analysis
Interpreting the changes in flow networks is important for understanding changes in mobility patterns and urban dynamics. Graphbased approaches were applied to quantify and compare the differences in the network structures. Firstly, two graph structures were constructed based on bike travel flow data for the two periods according to the method described above. Only trips with their origin and destination road segments in the 2,000 m buffers around metro stations were selected, as this was considered the maximum distance people would  Y. Yang, et al. Computers, Environment and Urban Systems 77 (2019) 101361 travel using the combination of dockless bike and metro. Table 4 shows the network properties in the new metro service catchments. From the table some changes between the 2 periods can be identified: the number of nodes increased from 984 to 1,064, and the number of edges increased from 34,948 to 41,967. Network connectivity, δ, representing graph structure resilience, also increased indicating the increased strength of the cycling network and its robustness to potential nodes failures, for example due to external disturbances such as a road closures. Node degree helps to evaluate the connectivity to and accessibility of destinations in a mobility graph (Zhong et al., 2014). Tables 5 and 6 confirm the significance of the changes in node degree mean and variance, with the mean value increasing from 124.4 to 132.5. Fig. 10 (a) shows the node degree distributions between the two periods and, despite higher means and third quantiles, the first quantile and median are relatively unchanged,    Alternative hypothesis is true: both true difference in means of degree and flux are not equal to 0.

Network statistical properties
Y. Yang, et al. Computers, Environment and Urban Systems 77 (2019) 101361 indicating the contribution of nodes with high degrees to the change in variance. This implies a higher probability of observing a node with very larger degree in the "after" period than "before". The node flux of graph, the total amount of trips that start from or end at a node, can be used to understand trip volumes. The rises in both total and mean flux (Tables 4 and 5) indicate that the demand for dockless bikes in local areas increased. The node flux CDF (Cumulative Density Function) in Fig. 10 (b) suggests there is higher probability of nodes with a larger flux. These changes describe the increased attractiveness of road segments, as well as their interaction strength in connection to other segments in the network. The node flux variability (Table 4) increases in value indicating a more heterogeneous distribution of interaction strength across the network. Tables 5 and 6 show the t-test and variance test results and confirm that observed changes are statistically significant. The clustering coefficient (Table 4) represents the extent to which nodes in a network tend to cluster (i.e. have links between them) (Saberi et al., 2018). Its average value shows an increase after the new metro opening, suggesting that the dockless bike trip network became more locally connected. These results suggest that the dockless bike sharing mobility network became denser and more heterogeneous after the opening of the new metro service. Similar patterns are evident in other graph properties. Firstly, changes in betweenness centrality were examined through "node betweenness" (Newman, 2005) rather than "edge betweenness" (Girvan & Newman, 2002). Node betweenness represents the extent to which nodes stand between each other in a graph, or serve as a bridge from one part to another (Newman, 2005). In the context of urban studies, this measure can identify hubs in flow networks (Zhong et al., 2014). Fig. 10 (c) shows the cumulative probability of node betweenness centrality. This indicates that the probability of low betweenness nodes decreased slightly, while the probability of higher betweenness nodes increased, and therefore that the nodes with higher levels of connectedness have a more intensive role in the "after" graph. This describes the emergence of well-connected hubs.
PageRank is another indicator graph node importance. Generally, if the number of highly centred nodes (with very high PageRank) deceases while the number of secondary PageRank nodes increases, then this implies a polycentric changes. These suggest that the influence of strong centred nodes has gradually relaxed with their centrality increasingly shared with emerging subcentres (Zhong et al., 2014). Fig. 10 (d) shows the CDF of PageRank, the "after" period exceeds the "before" period at around 0.75 of the cumulative probability and indicates that nodes are more likely to have secondary high PageRank, suggesting underlying polycentric transformation. The next section reinforces this finding by detecting community structures in graphs. Alternative hypothesis is true: both true ratio of variances of degree and flux are not equal to 1. Y. Yang, et al. Computers, Environment and Urban Systems 77 (2019)

Community detection and graph structure
Communities in a dockless bicycle sharing network can be understood as a set of road segment clusters (graph subsets, or sub-graphs) that are more connected by trips internal to the cluster than external. This different in internal and external connects in clusters is measured by modularity (Newman & Girvan, 2004). A modularity maximization algorithm, the Louvain method (Blondel, Guillaume, Lambiotte, & Lefebvre, 2008), was applied to detect the communities in the dockless bike network. Modularity characterizes the density of edges inside potential communities relative to edges outside of the community. Networks with high modularity have dense connections between community members but sparse connections with nodes in different communities. Optimizing this value theoretically results in the best possible segmentation of the nodes in a given graph structure. The Louvain method first finds small communities by optimizing modularity locally on all nodes, then each small community is grouped into one node and the first step is repeated, in an iterative way. Fig. 11 shows the different communities detected in the two periods using colour shades. Although the analysis does not include node (road segment) location, the communities are spatially coherent, suggesting considerable spatial structure in this network. The introduction of new metro station resulted in the emergence of new communities in the network (Fig. 11). One is located around the transfer station in the north, and another can be found near the centre of new network and demonstrates the impact of a new metro line on the structure of dockless cycling activities. Combined with the changes in other graph properties, it is possible to conclude that the new metro catchment area results in more bike trips and forms stronger local travel connections which are more polycentric.
Another important indicator graph structure is assortativity. This describes the tendency of high degree nodes (nodes with many edges), to connect to other high degree nodes. For example, the structures in Fig. 5 (d) and (f) have higher assortativity than the structures of Fig. 5 (c) and (e). Structures with high assortativity are more robust to node removal or failure. Assortativity increased from 0.1832 to 0.2895 after the introduction of the new metro service along with a slight increase in modularity, from 0.3236 to 0.3733. From a transportation perspective, the rise in the two indicators suggests that the new network alleviates congestion and enhances the efficiency of travel by non-motorised traffic (discussed in Sun, Zhang, & Wu, 2012). Here, the evolution in graph structure shows a quick self-adaptive process that meets the increasing and changing patterns of travel demand and traffic flows, and strongly implies underlying urban resilience. Whether the performance of the network will deteriorate with future increases in network size or a dramatic rise in trip numbers, remains to be seen. From the perspective of "urban metabolism", the assortative structure suggests a robust spread and interaction of various urban flows of information, capital and materials through non-motorised traffic, but may also pose future challenges, for example in the context of protecting against the spread of epidemics.

Flowmap
A final exploration of the changes in bike flow through flow maps is shown in Fig. 12. This illustrates bike travel origin or destinations road segment in the 2,000 m buffers and the colour shades are used to indicate trip density between different road segments, connected by the directed links. Several important differences can be observed, especially areas close to the new metro stations. For example, Cuiyuanlu, Xuefudadaodong, Wolongshan and Guotizhongxin stations all show large changes in flows and structures. These stations resulted in many new outward flows and regions of high trip destination density. The flows "i" and "ii" in Fig. 12 (a) suggest that bike users travel long distances from their origins to destinations in the "before" period, but these flows disappeared in the "after" period ( Fig. 12 b). This is because travellers used the metro service instead of taking long-distance cycling trips, implying local mobility improvements. In the before period, there are many travel flows on the east of Yingtanjie Station (a large residential area). After the opening of the new metro service, more bike travel flows extended from and linked to Yingtanjie station, indicating that residents quickly changed their travel habits and started to combine dockless bikes trips with metro travel.

Conclusion
In order to capture the impact of metro service on the dockless bike sharing system, this study applied a combination of geo-statistical and graph theory approaches. The analyses led to an in-depth understanding of the interaction and evolution of urban behaviours around non-motorised activity, by considering the changes in in the spatial patterns of dockless bike sharing as result of a new metro line. The new metro service increased nearby dockless bike sharing demand by 28%, and resulted in changes in other spatiotemporal patterns of travel behaviours, including bike travel distance and bike origin-destination spatial distributions. The observed changes in travel were not homogenous across the study area with greater impacts closer to new metro stations. The dynamics and evolution in the graph structure capture the urban resilience of dockless bike schemes who are able to adapt to infrastructural changes such as new metro systems. Dockless bike trip structure has a tendency towards being polycentric (i.e. with more community structures), stronger local connectivity, higher assortativity reflecting increased travel demand and scheme robustness. This insight is not only useful for dockless bike schemes, but also provides a new perspective for the analysis on urban resilience and the interdependence between different urban complex systems. Observations from flow maps can be used to indicate improvements in local mobility, and the speed of adaption by people to combine dockless bike for metro travels.
A number of policy and planning implications emerge from this study. First, it is important for bike sharing operators to prepare more bike fleets in the new metro service areas as demand will increase. Second, bike parking facilities need to be planned around metro stations over distances up to approximately 220 m. Third, analysis of the changes in the origin-destination and network structure can help to reveal which locations (roads) are more frequently used by bike users, thus supporting related planning. Last, bike fleet rebalance strategies need be redesigned to adapt to changes of flows.
There are several limitations to this work. First, it used short road segments (< 200 m) to aggregate bike travel flows, which although selected through analysis, may be specific to this study. Second, the study analysed data from a limited number of days. Climate (rainy days in summer) and the timing of the school year (the new term starts in September) were the major reasons for excluding data in order to minimise the variance of environmental factors. The changes in bike travel behaviours and associated flow structures could be confirmed if data for a longer period was available and through analysis of a similar case study area, with newly opened metro lines. Third, the data acquisition (section 3.2) procedure is only able to obtain information about the nearest 30 available bikes for each location, at each query. This potential data loss, if more bikes are available in that small area, introduces uncertainty into the analysis and results. Such shortcoming are commonly found in dockless bike sharing research (e.g. Ai et al., 2018;Liu et al., 2018;Shen, Zhang, & Zhao, 2018). However, since the Y. Yang, et al. Computers, Environment and Urban Systems 77 (2019) 101361 data are consistent over the two periods, with similar sampling characteristics and uncertainties, the findings from "before" and "after" the new metro line are comparable. Finally, although dockless bike data provide a better representation of non-motorised travel behaviour than cell phone data, there is an inherent sampling bias because not all city inhabitants will be cyclists or will use a bike sharing scheme. Future work will examine these issues further and will extend the analysis to integrate urban context, in order to develop a deeper understanding the implications of changes in spatiotemporal patterns and graph structure evolutions. It will also examine the long-term spatiotemporal effects of new metro stations on dockless bike sharing whose future patterns may change over time, as a result of deeper intergration of the two travel systems.