Transportation Research Part A 119 (2019) 170–180

Contents lists available at ScienceDirect

Transportation Research Part A

journal homepage: www.elsevier.com/locate/tra

The evolutionary interaction between taxi-sharing behaviours and

T
social networks
⁎
Yaoli Wang , Ronny Kutadinata, Stephan Winter
Infrastructure Engineering, University of Melbourne, Australia

A R T IC LE I N F O ABS TRA CT

Keywords: Mobility services increasingly consider the sharing of vehicles to reduce resource use and con-
Social network gestion. People are, however, reluctant to sharing vehicles with strangers. Therefore, this paper
Taxi-sharing investigates the dynamic co-evolution of taxi-sharing behaviours and the taxi-sharing-oriented
Travel behaviours social network structure. A social network based taxi-sharing method is introduced that is able to
Network evolution
prioritise taxi-sharing with acquaintances over strangers while capping the detour cost to rea-
Agent-based simulation
sonable and varied limits. Furthermore, the social network structure evolves and is updated
based on shared trips. An empirical simulation demonstrates the advantages of social network
based taxi-sharing, i.e., an increased match rate and a comparable satisfaction level to trip-based
methods. The spatial aggregation of the emerging social network not only suggests a space-time
searching heuristic for taxi-sharing, but also indicates how social factors conquer space while
space constrains social interactions.

1. Introduction

The concern on traffic congestion, environment protection, and time consumption in travel makes it necessary to explore greener
travel modes. Ridesharing is an emerging mode that combines real-time travel requests into the same vehicle to save traffic load on
the street. This mode is more tailored to ad hoc travel demands than general public transits and is supposed to be greener than private
vehicles. However, traditionally ridesharing matching and routing algorithms focus on computational, operational or service effi-
ciency. Few solutions for ridesharing were acknowledging social and behavioural issues until very recently, but only to a limited
degree. The consideration of trust in ridesharing has been mentioned in many studies as a significant barrier that prevents people from
accepting ridesharing (Amey, 2010; Chaube et al., 2010; Wessels, 2009; Santi et al., 2014; Gargiulo et al., 2015). A few blogs1,2 point
out that people subjectively dislike ridesharing because they do not know whether the people they will share rides with are pleasant,
nor are they willing to make detours for some strangers. In this regard, trust is discussed in terms of security, privacy, and comfort of
sharing rides with others.
One way to bring trust into the ridesharing system is to leverage social networks. A social network is a network made of individual
agents, and their dyadic ties representing some form of interaction. Common social networks are the virtual social networks where
people are connected on social media, or the real-life connections with people who are friends, family, and colleagues. Usually an
important factor of a social network connection is the trust between agents that is being built up throughout time, and thus social
networks are dynamic. The surveys by Chaube et al. (2010) and Wessels (2009) have provided concrete evidence that people have

⁎
Corresponding author.
E-mail addresses: wang.yaoli@yahoo.com (Y. Wang), ronny.kutadinata@arrb.com.au (R. Kutadinata), winter@unimelb.edu.au (S. Winter).
1
Why everyone hates UberPOOL. https://motherboard.vice.com/en_us/article/why-drivers-and-riders-hate-uberpool-and-lyft-line.
2
7 reasons why I hate Uberpool and Lyftline. https://therideshareguy.com/7-reasons-why-i-hate-uberpool-and-lyftline/.

https://doi.org/10.1016/j.tra.2018.10.043
Received 9 May 2017; Received in revised form 13 June 2018; Accepted 30 October 2018
Available online 29 November 2018
0965-8564/ © 2018 Published by Elsevier Ltd.
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

significantly higher willingness and detour tolerance to share rides with whom they know. That is because generally people have
higher trust in whom they know. In this sense, a social network in a ridesharing context is a trust system for the interaction in ride-
sharing. This social network is one of many social networks in daily life, but can be different for the reasons to be discussed later. This
paper studies ridesharing social networks, and for simplicity, calls them social networks. The connections in the trust network for
ridesharing is called friends hereafter.
This work uncovers from a theoretical perspective of a bi-directional process: how social networks are self-optimised and emerge
from ridesharing outcomes throughout time, which in turn affects the following ridesharings. Understanding the co-evolution of
ridesharing and its interdependent social network helps to address people’s incentive of commitment to new travel modes. Important
theoretical indications for policy makers are suggested why bringing social networks for the evaluation of trust into ridesharing is
crucial.
In a ridesharing system, Wang et al. (2017) have demonstrated that sharing rides exclusively with friends of the daily life social
networks (e.g., friends, colleagues) is not necessarily feasible, depending on the spatio-temporal distribution of the travel demands
and the friendship distribution. However, the trust in ridesharing links evolves over time: the strangers who share a couple of times
the same routes in their daily commuting can gain trust and be connected in a ridesharing social network. Travel behaviours, in
general, respond to different social engagements heterogeneously because of many reasons, such as varied utilities/preferences with
different affinities, or social network topological and spatial structures, to name a few (Hackney and Marchal, 2011; Carrasco and
Miller, 2009; Han et al., 2011). On the other hand, social networks will emerge from travel behaviours due to the co-presence of
activities (Hackney and Marchal, 2011). In ridesharing particularly, social networks evolve in both topological structure and spatial
distribution to overcome the trust barrier. New links grow out of trusted repetitive ridesharing trips, while infeasible links dissolve.
There are a few ridesharing products on the market that integrate social networks into ridesharing. For example, Faxi3 allows
travellers to form a social group or invite friends, and to see where friends are going so that they can choose to join them. The Chinese
ridesharing application Dida Chuxing4 lets travellers follow the people whom they have travelled with for future trips. That said, a
theoretical explanation and systematic examination on the dynamic co-evolution between social networks (in ridesharing) and ri-
desharing behaviours is still lacking.
In this study, taxi-sharing is investigated as an example. Taxi-sharing, as a particular type of ridesharing, is a process to match
(part of) passengers’ trips so that they travel in the same car, which is different from ridesharing that also considers drivers’ mobility
demands. The focus on taxi-sharing rather than the general ridesharing avoids the distraction of assigning the role of driver or
passenger to each traveller.
In addition to space-time constraints, many more factors may contribute to the co-evolution of social networks and taxi-sharing
behaviours as long as they add easiness to the process. For instance, an existing social network in daily life already reflects the trust
level at the moment, which can be used as the initial social network (S0 in Section 3.2) in ridesharing. The home locations of the
travellers affect people’s willingness to share rides. The traveller feedback after sharing a ride might have positive or negative input to
the trust level. However, this paper assumes that sharing a ride will always increase the trust level as this paper focuses on the spatial
aspect of a taxi-sharing-oriented social network along with its spatial evolution. The reason of the spatial focus is that taxi-sharing is
essentially a space-time bounded behaviour. Trust can be built up via repetitive travels if space and time allows. Additionally, ratings
can be integrated into the system easily once the data is available.
The research hypothesis is that, throughout a dynamic co-evolution process, a taxi-sharing method with social network preference
is advantageous over those ignoring social ties, yielding significantly higher match rates and suggesting an efficient search heuristic.
In order to prove this hypothesis, an agent-based simulation is built to investigate the evolution between taxi-sharing behaviours
and social network. To clarify, the resulted taxi-sharing-induced social network is a collaboration network particularly for taxi-
sharing, rather than a social network in a general sense. The initial condition of the social network can be adapted from a real social
network of any type, as long as higher trust is given to the connections. Social networks play a role in taxi-sharing in two stages: (1)
for finding potential partners, people grant higher detour tolerances to social contacts (called friends hereafter); and (2) the matching
optimisation by linear programming integrates varied travel satisfaction rates with persons of different affinities. The computed taxi-
sharing results then update the satisfaction rate with each person in the network due to their taxi-sharing experiences. The sa-
tisfaction rate is approximated by a utility function composed of factors such as the distance between their home locations, the
current degree of separation, the number of rides shared between them, the average detour cost of shared rides, and a no-match
penalty.
This work develops insights into the cooperative bundled travel behaviours in the light of not only spatio-temporal information
but also space-dependent social preferences. It proves the significance of a taxi-sharing algorithm prioritising social network pre-
ferences that leads to higher match rates. Heuristics are learnt from the spatial tendency of the social network evolution that disclose
to what extent social factors overcome space.
The rest of the paper is structured as follows. Section 2 summarises the state-of-the-art of ridesharing and geographic social
network analysis. Section 3 introduces the design of the simulation along with a utility function and a matching algorithm, followed
by its implementation and deployment in an experiment in Section 4. Section 5 collects the simulation results, and Section 6 discusses
these results. Section 7 concludes with the major findings and future work.

3
https://faxi.co.uk/.
4
http://www.didachuxing.com/.

171
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

2. The role and modelling of social networks in taxi-sharing

Studies have identified several factors that hold back people from taxi-sharing (and ridesharing in general). Amey (2010) sum-
marised four perspectives that hinder ridesharing: social, economic, institutional, and technical. The social aspect refers to the trust
between travellers, including the reluctance to share with strangers called “stranger danger”, the unreliability of service in case of an
emergency of a shared traveller, and the lower schedule flexibility that are essentially decided by travellers’ space-time budgets and
detour tolerance. Chaube et al. (2010) has demonstrated empirical evidence that 98% of people are happy to accept rides from a
friend, 69% from a friend of friend, but only 7% will accept rides from a stranger. Wessels (2009) found the significant difference in
detour tolerance descending with less familiar social acquaintances. Given the upper limit of the commuting time tolerance (He et al.,
2016), the resistance of switching to ridesharing is an obstacle to the scaling-up process of these systems (Kelly, 2016; Sulopuisto,
2016). The acceptance of shared autonomous vehicles is even more ambiguous considering the comfortability of sitting next to a
stranger in a limited space, without the presence of a typically trusted taxi driver (Krueger et al., 2016). Therefore, trust measure is a
key issue for autonomous taxi-sharing.
Hillier (2007) coined reflexive level activity for graph-induced behaviours while non-reflexive level activity for graph generating
activity. Social networks play a role on taxi-sharing in two reflexive ways: (1) Participation willingness and detour tolerance are
dependent on social acquaintance (Chaube et al., 2010; Wessels, 2009); (2) The geographic distributions of friendships and of the
trips each day delineate the physical potential of taxi-sharing. Therefore, this study investigates social networks from the perspectives
not only of topological structure but also of spatial distribution. On the other hand, the non-reflexive level activity is the feedback from
taxi-sharing experience on generating new social links.
The initialisation of social networks is grounded on the studies of the relationship between social networks and their spatial
embedment. Spatial clustering is a commonly observed phenomenon. Onnela et al. (2011) concluded that only small social groups are
geographically very tight, while larger groups split into clustered smaller ones over space. Gao et al. (2013) detected spatially
clustered community patterns in cell phone call networks. Shi et al. (2016), also with cell phone records, substantiated the existence
of a distance decay in social networks. Wong et al. (2006) demonstrated that generating social network links based on a distance
decay rule can capture the characteristics of small world networks. This work also investigates whether spatial aggregation exists in
the emerging taxi-sharing-oriented social network.
Furthermore, the construction of potential social network links can be inferred by people’s spatial travel history. Crandall et al.
(2010) built a probabilistic model and found co-location due to friendship is about 5,000 times by random chance. Cranshaw et al.
(2010) confirmed a positive correlation between the number of social ties and the location entropy, a measure on the diversity of a
place to which how many different people have visited. Correspondingly, Wang et al. (2015) found significantly more movement
overlaps between social network contacts versus random encounters. Zheng et al. (2011) proposed a model to recommend friends
based on the accordance of a sequence of visited locations with shared time slots, weighed by the popularity of locations. Pham et al.
(2013), Pham et al. (2016) did not only infer the existence of social ties but also proposed methods to infer social tie strengths.
However, these researches focus on the social network of joint activities (locations), which is orthogonal to a social network emerging
from bundled travels (routes). The present work is interested in the emerging social network from taxi-sharing over time, and
therefore only considers distance decay as a factor in social network evolution.

3. Methodology

This section first introduces a social network based taxi-sharing algorithm and optimisation (Section 3.1), followed by details for
the social network construction (Section 3.2), and finally looks into the dynamic mechanism between taxi-sharing and its emerging
social network from travel experiences (Section 3.3). Table 1 lists the variables frequently referred in this section.

3.1. Social network based taxi-sharing algorithm and optimisation

A known influence of social network on taxi-sharing is that people allow for heterogeneous detour tolerances due to varied social
affinities (Chaube et al., 2010; Wessels, 2009). A social network based taxi-sharing method is proposed that integrates heterogeneous
detour tolerance and satisfaction quantified by a symmetrical utility based on historical taxi-sharing experiences between a pair of
persons. On each day, the matching does a static pre-planning model alike to the algorithm in Wang et al. (2016) except it does not
consider alternative destinations but involves social networks. A match decides the sequence of (for simplicity) two trips’ pick-ups and

Table 1
Variable catalogue.
Variable Meaning Variable Meaning Variable Meaning

P Population pi Person i p Person i’s trips on day t

τt i
Xt Trips on day t xt , a a-th trip on day t O (xt , a) Owner of a trip
Ut Day t’s potential matches u (xt , a , xt , b) One potential match in Ut Mt Final matches on day t (from Ut )
p ,p Final matches between pi , pj δt (pi , pj ) Degree of separation on day t St Social network on day t
Mt i j
ct (pi , pj ) Utility of pi , pj Gt Complete graph for utilities in P

172
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

drop-offs subject to both travellers’ space-time budgets, assuming no traveller walks and there are sufficient vehicles (not explicitly
modelled) serving travel demands. If a travel fails to be matched with anyone’s, the traveller will travel alone.
p
Given a population P with a certain geographic distribution of their homes, each person pi ∈ P has a set of full day trips τt i for day
pi
t. A trip regardless of its owner is x t , a , where a is the index. All trips on day t form a full trip set Xt = {x t , a} = ∪i τt . Let O: Xt → P be the
function to find the owner of a trip. The matching involves four stages (Wang et al., 2016): (1) reading the whole day plan and
initialising each person’s travel agenda; (2) assigning travel budget (including detour tolerance) for each trip; (3) matching every each
pair of trips x t , a , x t , b ∈ Xt , (O (x t , a) ≠ O (x t , b)) , and finding all potential (satisfying both travellers’ space-time budgets) pairs of
matches u (x t , a , x t , b) ∈ Ut as the pre-computation matrix; (4) running binary integer programming weighed by utility to get the final
solution Mt .
Given the travel agenda Xt on day t, each trip x t , a ∈ Xt is associated with the earliest starting time EST (x t , a) , latest ending time
LET (x t , a), direct (shortest) travel time dTTime (xt , a) , and maximum detour budget maxDetTrip (x t , a) . The maxDetTrip (x t , a) is calculated
as min (dTTime (x t , a) ∗ DetRate (xt , a , xt , b), LET (x t , a) − EST (x t , a) − dTTime (x t , a)) , which is the maximum travel time for detour sa-
tisfying both psychological willingness and physical travel budget. DetRate (x t , a , x t , b) (1) is the percentage of a trip’s shortest travel
time the person would like to sacrifice depending on the degree of separation between the two travellers, δt (O (x t , a), O (x t , b)).

⎧30%, δt (O (x t , a), O (x t , b)) = 1,

DetRate (x t , a , x t , b) = 25%, δt (O (x t , a), O (x t , b)) = 2,
⎨
⎩10%, δt (O (x t , a), O (x t , b)) = ∞. (1)
Stage 4 maximises the number of matches. Each u (x t , a , x t , b) ∈ Ut is binary: 1 (matched) or 0 (not matched), subject to that one trip
can be matched to up to one other trip. The matrix is weighed by ct (O (x t , a), O (x t , b)) , the utility between two travellers. Multiple
matches between the same pair of persons on day t are assigned the same weight. The final output from binary linear programming is
Mt = {(x t , a , x t , b) | u (x t , a , x t , b) = 1} . The matched trips between (pi , pj ) is defined as
pi , pj
Mt = {(x t , a , x t , b) | (x t , a , x t , b) ∈ Mt ;O (x t , a), O (x t , b) ∈ {pi , pj }} . Hence, the total number of matches up to day t − 1 is:
t−1
p ,p p ,p
nt ′ i< jt = ∑ |Mt ′ i j |.
t ′= 1 (2)
p ,p p ,p
The maximum value of nt ′ i< jt , used in Section 3.3 to normalise nt ′ i< jt , is approximated by:
t−1
p ,p p p
Nt ′ i< jt = ∑ min { |τt ′ i |, |τt ′ j |},
t ′= 1 (3)
p
where |τt ′ i |
= |{x t ′, a | O (x t ′, a) = pi }|. Theoretically, the maximum number of matches cannot be larger than any of the travellers’
number of trips per day. Let the actual detour cost of a match (x t , a , x t , b) ∈ Mt be detCost (x t , a , x t , b) , simply quantified by the average
percentage of shortest travel time of x t , a , x t , b for detour, the total detour cost of (pi , pj ) until day t is:
p ,p
detCostt ′ i< jt = ∑ detCost (x t ′, a , x t ′, b).
p , pj
{ ( x t′, a, x t′, b ) ∈ Mt′ i | 1 ⩽ t ′< t } (4)

3.2. Construction of taxi-sharing-oriented social networks

On day t, the social network St (t ⩾ 1) encapsulates all pairs of persons who are directly connected, i.e., degree of separation
δt (pi , pj ) = 1. If two persons are only indirectly connected via a third person, δt (pi , pj ) = 2 ; otherwise they are strangers δt (pi , pj ) = ∞.
The initialisation of a social network, S0 , is to permutate links with δ0 (pi , pj ) = 1 in P.
A complete graph Gt = < P , E , Ct > is built to represent utility between each pair of persons, where ct (pi , pj ) ∈ Ct is the utility for
the pair e (pi , pj ) ∈ E . A direct social network link will be built if ct (pi , pj ) goes over a threshold SocThres = w h + 0.6·w fr , i.e.,
δt (pi , pj ) = 1 ⇔ ct (pi , pj ) > SocThres ; if the utility drops below this threshold, the link will be dissolved. The threshold is built so that
no link can be upgraded to 1st degree without any taxi-sharing history. Since 0.6 is the friendship utility of 2nd degree friends (to be
specified in (6b)), and 1 is the maximum utility of home location, SocThres guarantees that even if a 2nd degree link with the highest
utility of home location cannot evolve into a 1st degree link unless taxi-sharing history adds up to it.

3.3. The dynamic co-evolution of taxi-sharing and social network

The dynamic co-evolution of taxi-sharing and its induced social network goes in such a procedure (Fig. 1): (1) update the utility to
ct (pi , pj ) for each e (pi , pj ) ∈ E ; (2) update the social network from St − 1 to St based on the latest utility; (3) run taxi-sharing simulation
(Section 3.1) for Xt ; (4) calculate Mt and get the most up-to-date ridesharing history Ht by appending Mt to Ht − 1.
Travellers’ preferences are based on the following premises backed up by previous studies:

• The proximity of origins counts more than that of destinations (Vanoutrive et al., 2012), which explains why home distance
contributes to travel utility.
• Detour tolerance increases with social acquaintance (Chaube et al., 2010; Wessels, 2009), which reflects the trust issue in taxi-
173
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

Fig. 1. The dynamic interaction between taxi-sharing and social network.

sharing.
• The collaboration network structure is affected by the initial social network structure that can be of any type as long as it gives a
person higher trust to the connected persons.
• Trust and preference are gained or lost from taxi-sharing experiences, including:
(1) People prioritise lower detour cost;
(2) The more the times two persons share rides, the higher the trust;
(3) The above two rewards are subject to diminishing marginal utility (Kauder, 1965);
(4) Lacking sharing experience reduces travel utility and thus dissolves collaboration links.

Therefore, the travel utility encompasses three factors: the distance between riders’ homes that affects their possibility to share
rides, current social network status, and taxi-sharing history (Fig. 1). Taxi-sharing history up to date includes the average detour cost
of past rides, the number of matches until yesterday, and the no-match penalty. As one feasible function that incorporates the
aforementioned factors, the utility function for e (pi , pj ) is formulated as:

ct (pi , pj ) = w h·c h (·,·) + w fr ·ct −fr 1 (·,·) + w md·c m d −n −n

< t (·,·)· c < t (·,·) + w · ct (·,·) (5)
In (5), c h (pi ,
pj ) is the utility from home distance d (pi , pj ) of two persons, ct −fr 1 (pi ,
pj ) is the utility based on current friendship,
−n
cm d
< t (pi , pj ) is the utility from the number of matches in history, c < t (pi , pj ) is the utility from detour cost in history, and ct (pi , pj ) , a
negative value, is the penalty on friendship for no-match. w h, w fr , w md ,and w−n are the weights for each term. To show the overall
ride experience throughout time rather than of one ride, the utilities from number of matches and detour cost have a joint function
and weight.
The specific formulas for each term is defined below (t ⩾ 1):
c h (pi , pj ) = fh (d (pi , pj )) (6a)

⎧ 0.9, δt (pi , pj ) = 1
⎪
ct −fr 1 (pi , pj ) = 0.6, δt (pi , pj ) = 2
⎨
⎪ 0.1, δt (pi , pj ) = ∞. (6b)
⎩
p ,p
ln(nt ′ i< jt + 1)
cm
< t (pi , pj ) = p ,p
ln(Nt ′ i< jt + 1) (6c)
p ,p
⎧1 − detCost t ′ i< jt
, t>1
c d< t (pi , pj ) = p ,p
nt ′ i< jt
⎨
⎩ 0, t=1 (6d)
☆
1 − e (t − t − 1)
ct−n (pi , pj ) =
‖1 − e (D − 1) ‖ (6e)
In (6a), the utility dependent on d (pi , pj ) , the distance between the two persons’ homes, can be specified in particular contexts (see
Section 4.3). Eq. (6b) indicates a higher utility (i.e., taxi-sharing priority) for closer friends. Eq. (6c) quantifies the influence of the

174
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

p ,p
total number of matches between a pair of persons in history nt ′ i< jt (2). The logarithmic function indicates the diminishing marginal
utility of each extra trip. The function is normalised with the utility value by plugging in the theoretically maximum number of
p ,p p ,p
matches Nt ′ i< jt (3). Eq. (6d) is the utility from the average detour cost of matches in history between pi , pj , where detCostt ′ i< jt is the
total detour cost specified in (4). Sub-utilities (6c) and (6d) are positive feedback from taxi-sharing experience. Eq. (6e), on contrast,
is the negative feedback from taxi-sharing that punishes a friendship with no-match. t ☆ ⩾ 0 is the last day when the pair had a match,
and D is the total duration of the simulation. The exponential form reflects stronger retainment of friends at the beginning, when
people still memorise their links even though they have not met for some time. The penalty grows faster, however, when two persons
do not have any contact for longer time.

4. Experiment design

The simulation is implemented in Repast Simphony (https://repast.github.io/repast_simphony.html) and tested in the shire of
Yarra Ranges, composed of several suburbans located east and northeast to Melbourne, Victoria. The parameter settings are listed in
Section 4.1. The simulation runs a period of 30 days (D = 30 ) with the processed datasets by Jain et al. (2017) based on the Victorian
Integrated Survey of Travel and Activity (VISTA) 2009–2010 (Victorian Department of Transport, 2011). Trip types include not only
commuting trips to work but also trips for daily and leisure activities such as shopping at a supermarket, going to the gym, or picking
up/ dropping off kids. The trips are assumed to be sampled for one weekday, since the surveys were conducted as a one-weekday
survey of each person, but different days of different persons (Victorian Department of Transport, 2011). To tailor for this study, a
trip-regeneration processing (Section 4.2) is used to permutate trips for 30 days. Section 4.3 discusses the spatial distribution of the
sampled trips along with its influence on the simulation. Section 4.4 details the implemented social network initialisation.

4.1. Simulation runs and parameters

In (5), the weight of each term in the utility function determines how crucial a role each term contributes. Higher ratios of w md to
other weights from non-sharing factors (e.g., w h, w fr ) mark higher influence of sharing experience on the consequential social net-
work structure, while distance decay and current social network structure play smaller parts. The weights for each term in the
simulation are designed as shown in Table 2. Setting#3 is the most impacted one by taxi-sharing. Each term in (5) is normalised
between 0 and 1 (inclusive), as formulated in (6).

4.2. Generating recursive trips over 30 days from VISTA

Given that VISTA data only samples one-weekday travels, trips of the 30 days are assumed to have a recursive period of 5 days
(excluding weekends). The first 5 days are permutated from the processed VISTA data by introducing the randomness to flexible trips
with either flexible travel time or location (defined in Wang et al., 2016) that may or may not occur on different days. With the
permuted 5 days’ trips as the baseline, the rest days will repeat the trips on their corresponding day, e.g., day 5 repeats day 0. The only
difference of recursive travels on different periods is the dynamically updated utility of taxi-sharing with a certain person, which
consequentially leads to varied matching outputs.

4.3. Home distribution and distance decay function

As part of the utility function (5), home distance contributes to the travel partner choice. The particular distance decay function is
defined as (7) because of the heterogeneity in terms of landscape and home distribution in Yarra Ranges. The average distance
between two spatial clusters of population is about 10 km. The step-function creates a heuristic that inclines to assign higher utility to
people within the same spatial cluster (⩽ 10 km) who have higher chances for taxi-sharing. Though home locations cannot fully
capture the characteristics of human movement behaviours, they are one of the feasible representatives since people’s activity spaces
are regularly bounded by such “anchor points” as home (González et al., 2008; Huang et al., 2010; Wang et al., 2015).

⎧ 0.1, d (pi , pj ) > 10

fh (d (pi , pj )) = −β
⎨ α·(d (pi , pj ) + 1) , d (pi , pj ) ⩽ 10
⎩ (7)

When plugging in the values that d (pi , pj ) = 0, fh (d (pi , pj )) = 1, and d (pi , pj ) = 10, fh (d (pi , pj )) = 0.1, we get α = 1, β = log1110 .

Table 2
Weights of terms in utility function.
Setting# wh w fr w md w−n

1 1
2 0.2 1 2 2
3 10

175
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

4.4. Initialisation of social networks

Two types of social network initialisations are designed to compare their corresponding taxi-sharing outputs: (1) a null model
(from scratch) that assigns no 1st degree link at the beginning, and observes the emerging social network from taxi-sharing experience;
and (2) a small world model that is topologically locally well-connected but with relatively short average distance between nodes,
which is believed to capture the topological structure of real social networks (Milgram, 1967; Amaral et al., 2000; Watts and Strogatz,
1998). The initialisation of small world social network is generated with the WattsBetaSmallWorldGenerator module in Repast. Despite
of no universally recognised network density, the average direct contacts per person is set to 6 in the population of 714 people,
yielding a low density of 0.84%, which is widely observed in real world networks (Faust, 2006). The pair of persons with an initialised
link is set with δt (pi , pj ) = 1. In null model cases, no link is generated, and therefore, all δt (pi , pj ) = ∞, ( ∀ i, j;i ≠ j ) at the beginning.
Any 2nd degree link is derived from the 1st degree links.

5. Results

The dynamics of four variables are tracked over time (Fig. 2): (1) the average number (degree) of direct friends per person (2a);
(2) the average distance between homes of direct friends (2b); (3) the number of matches on each day over the period (2c); and (4)
the fluctuation of the daily average utility of direct friends throughout time (2d). In each figure’s legend, “1”, “2”, “3” represent
Setting# in Table 2, and “sw”, “null” are for small world networks and null networks, respectively.
a. Degree distribution. In null network models, the average degree steadily grows to a stable plain (from 0 to 0.084, 0.232, and
1.188 for each setting). A similar trend (with a shift from the origin) appears at the early and middle stages in small world models
until a sharp drop erupts on day24 resulting from the broken links due to no-match penalty (6e). While real social network links do
not break down easily, the simulation shows only the fast-forward process of the collaboration links representing the feasibility of taxi-
sharing partnership. Small world models eventually converge to a slightly higher degree (0.26, 0.487, and 1.384 for each setting)
than their counterparts in null models at equilibrium.
b. Spatial aggregation. The average distance between direct friends’ homes in small world models is decreasing over time. On
Fig. 2b, an obvious nadir of the distance pops up on day24 when unused links break down, followed by a bounce-back to the final
equilibrium. Results from different weights of taxi-sharing experience are not alike. The more the experience weighs in (Setting#3 to
Setting#1), the faster and lower the average home distance drops before the nadir. A similar order occurs at equilibrium as well. At
the nadir, however, the rank is the opposite. On comparison, the average home distance from null models is smooth, slightly in-
creasing to the equilibrium over time except a tiny hump at the beginning of each curve due to the scarcity of matches and ran-
domness in space at the beginning. Nevertheless, Setting#1 from neither model yields distinguishable variation of average home
distance owing to its low sensitivity to the reward of taxi-sharing.
c. Number of matches. The number of matches per day displays a recursive pattern as an inherent result from the repeated travel
pattern of 5 days, but is not obviously related to the degree of friends, the average home distance, or the average utility (Figs. 2a–c).
Nevertheless, the daily number of matches from small world models is higher than their counterparts in null models. Higher weight of
taxi-sharing also leads to higher match rate.
d. Average utility. The average utility between direct friends in small world models increases a little in the first few days and
stays steady until day24 when a distinguishable rise appears. The sharp rise is consistent with the break-down of the unused links, a
self-optimising process to maximise utility. Though the results from the null models demonstrate the opposite trend, decreasing
steadily all the way, they converge to a statistically identical utility ( p < 0.05) to the small world models at the equilibrium.

6. Discussions

Indicated from the simulation results, social-induced taxi-sharing and the social (collaboration) network mutually reinforce each
other in favour of future taxi-sharing outcomes. This is a reflection of the deeper level phenomenon that society is driven both by
socio-economic factors and by spatial laws (e.g., distance decay) (Hillier, 2007; Batty, 2008). The major points are argued as below:
The equilibrium proves the existence of the stable status accommodating social factors and geographical constraints.
Outcomes from simulations, no matter considering or ignoring social networks, all demonstrate an equilibrium in each case. Fig. 2b
shows that the consequential networks at the equilibrium from small-world social network initialisations are more spread over in
space, which means higher detour cost sharing taxis with friends. However, the corresponding simulations with and without initial
social networks converge to statistically identical average levels of satisfaction (i.e., utilities) with the rides (Fig. 2d after day24). That
is to say, higher detour cost with friends does not affect the eventual satisfaction level of travel, while it does increase the number of
matches in total (see next paragraph). This is a consequence of balancing detour cost and travel partner preference. On one hand, for
the cases with initial social networks, even if the matches prioritising friendships are not always of the lowest detour cost, they are
traded off by trust. On the other, when no initial social networks exist, low trust is balanced off by lower cost shares with strangers
nearby. The indication for industries is to introduce the trust network (i.e., social network in ridesharing) at an early stage so that the
system bootstraps itself quickly for more matches. The equilibrium is assumedly subject to the specific context of geography and
population, which means more empirical studies of different contexts are needed when data is available.
Incorporating social networks into taxi-sharing can lead to a higher number of matches than ignoring social networks. In
contrast to identical satisfaction levels and higher detour cost, the outputs with initial small world networks consistently yield

176
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

Fig. 2. Output results from taxi-sharing over 30 days (coloured). (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)

significantly higher numbers of matches everyday than their corresponding results from null models (Fig. 2c). The nonnull social
network (i.e., small-world network) models embark the simulation with a higher level of trust, namely, higher detour tolerance for
friends. Over the period of 30 days, the robustness of some initialised links is reinforced by collaborations. Discernible on Fig. 3, the
yellow links, which are the retained links in small world on day29, exist between some relatively long-distance pairs. This is par-
ticularly obvious when social network weighs in more, i.e., in Setting#1 and Setting#2 (the first and second graphs in Fig. 3). The
null models that treat everyone as strangers, on contrast, are more conservative on detour tolerance from the beginning, yielding
fewer matches (Figs. 2c). Taxi-sharing granting higher tolerance to friends beneficially enhances the number of matches, which

177
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

Fig. 3. Distributions of social networks on day0, day24, and day29 (coloured). (For interpretation of the references to colour in this figure legend,
the reader is referred to the web version of this article.)

indicates how social network conquers space. The significance of this work hence is to make a theoretical argument that a ridesharing
system should take into account the influence of the trust network from the very beginning.
Local strangers substitute remote friends to optimise the collaborative social network structure. The existing spatially
random social network (i.e., the small world network initialised without spatial distribution regulation) is observed growing spatially
aggregated over time. Taxi-sharing is a type of spatially restricted social interactions that require co-presence. This is opposite to the
observation by Hackney and Marchal (2011) that social networks are growing in space when people travel for spatially free social
events. The extent of restriction is partially ascribed to how heterogeneous the landscape is. Melbourne built-up area is assumed to
develop a spatially more spread social network from strangers than Yarra Ranges with many mountains and rivers. Spatially local
taxi-sharing drives social link production by spatio-temporal co-presence (constraint), particularly, by overlapped travels. Therefore,
a social network based algorithm not only reserves the existing social links sufficiently strong though of longer distance (while

178
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

eliminating those long-distance unused links); but it also develops local short-distance links with those who were originally strangers
but are proved efficient and trustworthy during taxi-sharing experience. The observation substantiates the effectiveness of designing
ridesharing systems (e.g., Faxi and Dida Chuxing) where people can follow and choose to join their previous travel partners in future
requests.
Back to the hypothesis, the social network based taxi-sharing is advantageous. Social network based taxi-sharing creates a
heuristic filter that integrates the benefit of spatial proximity and the priority to friends. As aforesaid, instead of causing the loss of
getting rides, preferring friends can even increase match rates. Moreover, feasible local links are constructed throughout the self-
optimisation of the collaboration network. The heuristic, therefore, is that searching strangers for a match should be indexed by
spatial proximity to reduce time consumption, while searching friends by social links satisfies travel preference.

7. Conclusions and future work

This work investigates the interactive dynamics between people’s taxi-sharing behaviours and a taxi-sharing-oriented social
network structure. The simulation is conducted with a suggested social network based taxi-sharing method that infuses varied utilities
(rather than friend-only) for taxi-sharing with different partners into an existing taxi-sharing algorithm. The simulation is based on
simplified assumptions only to lay the foundation for further studies into spatio-social systems. Getting back to the hypothesis, a
social network based taxi-sharing method is superior to normal sharing methods because: (1) It increases the overall match rate while
still keeping an identical satisfaction level of travel; (2) It utilises the attraction of social preference and trust to overcome the
drawback of geographic barriers, which optimises the collaboration network structure; (3) It indicates a searching heuristic that
feasible matches can be retrieved by both spatial and social indices. Industries in ridesharing sector should leverage the findings in
order to achieve better product designs.
A more pragmatic and refined investigation into the dynamic co-evolution should takes into account the positive and negative
feedback in the rating system. According to Kooti et al. (2017), the rating is dependent on factors such as age, gender, and the
percentage of earning from surge pricing. Such observation makes it difficult to reasonably simulate the rating distribution without
any access to realistic dataset. That said, it is worthwhile to seek for available data of realistic behaviours and re-investigate the
results. As a generic category of algorithms, social network based taxi-sharing can be integrated into any other taxi-sharing or
ridesharing algorithm. The social network also works as a heuristic for faster retrieval of potential partners, which may partially solve
the calculation efficiency issues reported by Wang et al. (2016). Further studies can also release the constraint of matching only two
persons, and also explicitly model vehicles with a controlled fleet size. It is also worth examining a fully self-organised social-induced
taxi-sharing system where no centralised taxi-sharing optimisation is conducted and the utility between a pair of persons is not
mutual.

Acknowledgements

This research has been supported by the Australian Research Council (LP 120200130).

References

Amaral, L.A.N., Scala, A., Barthelemy, M., Stanley, H.E., 2000. Classes of behavior of small-world networks. Science 97 (v), 11149.
Amey, A.M., 2010. Real-time ridesharing: exploring the opportunities and challenges of designing a technology-based rideshare trial for the MIT community, Master's
thesis, Massachusett Institute of Technology.
Batty, M., 2008. The size, scale, and shape of cities. Science 319, 769–771.
Carrasco, J.A., Miller, E.J., 2009. The social dimension in action: a multilevel, personal networks model of social activity frequency between individuals. Transp. Res.
Part A: Pol. Pract. 43, 90–104.
Chaube, V., Kavanaugh, A.L., Pérez-Quiñones, M.A., 2010. Leveraging social networks to embed trust in rideshare programs. In: Proceedings of the 43rd Hawaii
International Conference on System Sciences. Honolulu, HI.
Crandall, D.J., Backstrom, L., Cosley, D., Suri, S., Huttenlocher, D., Kleinberg, J., 2010. Inferring social ties from geographic coincidences. Proc. Nat. Acad. Sci. 107
(52), 22436–22441.
Cranshaw, J., Toch, E., Hong, J., 2010. Bridging the gap between physical location and online social networks. In: Proceedings of the 12th ACM International
Conference on Ubiquitous Computing, pp. 119–128.
Faust, K., 2006. Comparing social networks: Size, density, and local structure. Metodološki zvezki 3 (2), 185–216.
Gao, S., Liu, Y., Wang, Y., Ma, X., 2013. Discovering spatial interaction communities from mobile phone data. Trans. GIS 17 (3), 463–481.
Gargiulo, E., Giannantonio, R., Guercio, E., Borean, C., Zenezini, G., 2015. Dynamic ride sharing service: are users ready to adopt it? Procedia Manuf. 3, 777–784.
González, M.C., Hidalgo, C.A., Barabási, A., 2008. Understanding individual human mobility patterns. Nature 453 (7196), 779–782.
Hackney, J., Marchal, F., 2011. A coupled multi-agent microsimulation of social interactions and transportation behavior. Transp. Res. Part A: Pol. Pract. 45 (4),
296–309.
Han, Q., Arentze, T., Timmermans, H., Janssens, D., Wets, G., 2011. The effects of social networks on choice set dynamics: results of numerical simulations using an
agent-based approach. Transp. Res. Part A: Pol. Pract. 45, 310–322.
He, M., Zhao, S., He, M., 2016. Tolerance threshold of commuting time: evidence from Kunming, China. J. Transp. Geogr. 57, 1–7.
Hillier, B., 2007. What do we have to add to a social network to get a society? Answer: something like what we have to add to a spatial network to get a city. In: Winter,
S., Robins, G. (Eds.), International Workshop on Social Space and Geographic Space, in conjunction with Conference on Spatial Information Theory 2007, pp.
35–50.
Huang, W., Dong, Z., Zhao, N., Tian, H., Song, G., Chen, G., Jiang, Y., Xie, K., 2010. Anchor points seeking of large urban crowd based on the mobile billing data. In:
International Conference on Advanced Data Mining and Applications. Springer, Berlin Heidelberg, pp. 346–357.
Jain, S., Ronald, N., Thompson, R., Winter, S., 2017. Predicting susceptibility to use demand responsive transport using demographic and trip characteristics of the
population. Travel Behaviour Soc. 6, 44–56.
Kauder, E., 1965. The law of diminishing utility. In: History of Marginal Utility Theory. Princeton University Press, pp. 135–142.

179
Y. Wang et al. Transportation Research Part A 119 (2019) 170–180

Kelly, É., 2016. Helsinki’s ambitious Uber for buses experiment has failed. What went wrong? URL <http://sciencebusiness.net/news/77416/Helsinki%25E2%2580%
2599s-ambitious-Uber-for-buses-experiment-has-failed.-What-went-wrong>.
Kooti, F., Grbovic, M., Aiello, L.M., Djuric, N., Radosavlejevic, V., Lerman, K., 2017. Analyzing Uber’s ride-sharing economy. In: International World Wide Web
Conference Committee, pp. 574–582.
Krueger, R., Rashidi, T.H., Rose, J.M., 2016. Preferences for shared autonomous vehicles. Transp. Res. Part C: Emerging Technol. 69, 343–355.
Milgram, S., 1967. The small world problem. Psychology Today 2, 60–67.
Onnela, J., Arbesman, S., González, M.C., Barabási, A., Christakis, N.A., 2011. Geographic constraints on social network groups. PLoS ONE 6 (4), e16939.
Pham, H., Shahabi, C., Liu, Y., 2013. EBM - An entropy-based model to infer social strength from spatiotemporal data. In: Proceedings of the 2013 ACM SIGMOD, pp.
265–276.
Pham, H., Shahabi, C., Liu, Y., 2016. Inferring social strength from spatiotemporal data. ACM Trans. Database Syst. 41 (1), 7:1–47.
Santi, P., Resta, G., Szell, M., Sobolevsky, S., Strogatz, S.H., Ratti, C., 2014. Quantifying the benefits of vehicle pooling with shareability networks. Proc. Nat. Acad. Sci.
111 (37), 13290–13294.
Shi, L., Wu, L., Chi, G., Liu, Y., 2016. Geographical impacts on social networks from perspectives of space and place: an empirical study using mobile phone data. J.
Geogr. Syst. 18 (4), 359–376.
Sulopuisto, O., 2016. Why Helsinki’s innovative on-demand bus service failed. URL <http://citiscope.org/story/2016/why-helsinkis-innovative-demand-bus-service-
failed>.
Vanoutrive, T., Van De Vijver, E., Van Malderen, L., Jourquin, B., Thomas, I., Verhetsel, A., Witlox, F., 2012. What determines carpooling to workplaces in Belgium:
location, organisation, or promotion? J. Transp. Geogr. 22, 77–86.
Victorian Department of Transport, 2011. Victorian integrated survey of travel and activity (2009–2010). URL <http://economicdevelopment.vic.gov.au/transport/
research-and-data/vista/vista-data-and-publications>.
Wang, Y., Kang, C., Bettencourt, L.M.A., Liu, Y., Andris, C., 2015. Linked activity spaces: Embedding social networks in urban space. In: Helbich, M., Arsanjani, J.J.,
Leitner, M. (Eds.), Computational Approaches for Urban Environments, pp. 313–336. (Chapter 13).
Wang, Y., Kutadinata, R., Winter, S., 2016. Activity-based ridesharing: Increasing flexibility by time geography. In: Proceedings of the 24th ACM SIGSPATIAL, pp.
1:1–1:10.
Wang, Y., Winter, S., Ronald, N., 2017. How much is trust: the cost and benefit of ridesharing with friends. Comput. Environ. Urban Syst. 65, 103–112.
Watts, D.J., Strogatz, S.H., 1998. Collective dynamics of ‘small-world’ networks. Nature 393, 440–442.
Wessels, R., 2009. Combining ridesharing & social networks. In: ITS World Congress. URL <https://www.utwente.nl/ctw/aida/education/ITS2-RW-Pooll.pdf>.
Wong, L.H., Pattison, P., Robins, G., 2006. A spatial model for social networks. Phys. A 360 (1), 99–120.
Zheng, Y., Zhang, L., Ma, Z., Xie, X., Ma, W., 2011. Recommending friends and locations based on individual location history. ACM Trans. Web 5 (1), 5:1–44.

180