Impact of Detour-Aware Policies on Maximizing Profit in Ridesharing

Arpita Biswas Ragavendran Gopalakrishnan Theja Tulabandhula

[email protected] [email protected] [email protected]
Indian Institute of Science Cornell University University of Illinois Chicago

Asmita Metrewar Koyel Mukherjee Raja Subramaniam Thangaraj

[email protected] [email protected] [email protected]
Xerox Research Center India IBM Research India Amazon

Abstract 1 Introduction
This paper provides efficient solutions to maximize Ridesharing is a key initiative that can curb ever-increasing
profit for commercial ridesharing services, under a congestion on the urban transportation network and is an im-
pricing model with detour-based discounts for pas- portant means to move the world towards a sustainable future.
sengers. We propose a greedy heuristic for real- While the term ridesharing has been used to include peer-to-
time ride matching and compare with other heuris- peer carpooling platforms as well, here we focus on commer-
tics. We study the trade-offs between optimality cial ridesharing service providers such as LyftLine and Uber-
and execution time that the solutions offer. Simula- Pool, that hold a significant share of the ridesharing popula-
tions on New York City (NYC) taxi trip data show tion.
that our heuristics are up to 93% optimal and 105 While ridesharing is undoubtedly appealing from a sus-
times faster than the exponential-time optimal al- tainability perspective, profit maximization is the primary
gorithm. goal of commercial providers. Given the emerging literature
Commercial ridesharing service providers generate on detour-aware pricing and routing policies, e.g., [Jung et
significant savings by matching multiple ride re- al., 2013; Gopalakrishnan et al., 2016] that enhance the ride
quests, which are typically shared between the ser- experience, algorithms for profitable ride-matching should
vice provider (increased profit) and the rideshar- adapt to keep up. Such quality-enhancing policies encour-
ing passengers (discounts). It is not clear apriori age increased adoption of ridesharing, but perhaps at the
how to split, since higher discounts would encour- cost of reduced average per-ride profit. We study three con-
age more ridesharing, thereby increasing total sav- nected issues centered around profit optimization in commer-
ings, but the fraction of savings taken as profit is cial ridesharing.
reduced. We simulate a scenario where the deci- First, it is unclear how to tune real-time ride-matching al-
sions of the passengers to opt for ridesharing de- gorithms to quickly find a near-optimal set of matches that
pend on the discount offered by the provider. We results in high profit under detour-based discounts (the exact
provide an adaptive learning algorithm IDFLA that optimization is an NP-hard problem [Cordeau et al., 2006]).
learns the optimal profit-maximizing discount fac- Second, ridesharing generates significant economic sav-
tor for the provider. An evaluation over NYC data ings by reducing the driver-miles necessary to serve a set
shows that IDFLA, on average, learns the optimal of passengers, which are split between the service provider
discount factor under 16 iterations. (their profit) and the passengers (as discounts). It is not obvi-
ous how this split should be effected, since higher discounts
Finally, we investigate the impact of imposing a encourage more ridesharing, increasing total savings, but the
detour-aware routing assuring sequential individual fraction of savings that constitutes the profit is smaller.
rationality which offers a better ride experience and Third, in order to survive in an increasingly competi-
increasing the provider’s market share, but at the tive market, ridesharing providers adopt additional quality-
cost of decreased average per-ride profit due to the enhancing features which do not align fully with the goal of
reduced number of matched rides. We construct a profit maximization, perhaps resulting in lower profit (due to
model that captures these opposing effects, wherein the smaller feasible set of quality-compliant rides). It is im-
simulations on NYC data show that 7% increase in portant to understand the overall impact on profitability, and
market share would suffice to offset the decreased ensure that any resulting increase in market share can counter
average per-ride profit. the reduced average per-ride profit.
Keywords: Ridesharing, Profit Maximization, Adaptive The issues mentioned above are crucial for commercial
Learning, Simulation, Heuristics ridesharing platforms; however, they are less explored. We
address these issues, give solutions, and provide extensive
This work was done when the authors were working with Xerox evaluation on a real-world dataset.
Research Center India (XRCI).
Our Contributions: We investigate three important aspects amount paid to the driver. When compared to assigning each
of profit optimization in commercial ridesharing, resulting in user their own cab, ridesharing requires fewer cabs to serve
the following contributions: the same set of users, leading to increased profit.
• We provide an Integer Linear Program (ILP), and also an At the core of these platforms is a real-time matching algo-
efficient greedy heuristic that quickly matches passen- rithm that runs at regular time intervals, and generates a set of
gers to rides to approximately maximizes profit under a matched users from a set of waiting users (preprocessed to ac-
detour-based discount policy. We evaluate their perfor- commodate other spatio-temporal constraints). Such a system
mance experimentally (Section 3). would maintain a pool of pending requests, and then invoke
the matching algorithm every few minutes to determine op-
• Under a model in which the likelihood of opting for timal matches. We provide a matching heuristic that directly
ridesharing depends on the offered discount, we propose optimizes profit, under a detour-based discount scheme. Any
an adaptive algorithm that learns the profit-maximizing unmatched requests left at the end of the algorithm either wait
discount parameter (Section 4). for another round, or, after a certain waiting time threshold
• We additionally impose a detour-aware routing policy elapses, are assigned fresh, empty cabs. Thus, it is important
based on sequential individual rationality, introduced to provide a time-efficient matching algorithm for the larger
in [Gopalakrishnan et al., 2016]. Using simulations, dynamic ridesharing system to be sufficiently responsive to
we investigate the minimum increase in market share incoming user requests.
needed to offset the reduced average per-ride profit (Sec- Moreover, under a detour-based discount policy, shared
tion 5). rides with smaller detours are more likely to be profitable.
However, such a myopic observation could lead to subopti-
2 Related Work mal solutions, thus rendering the problem nontrivial.
In Section 3.1, we provide an Integer Linear Program (ILP)
There is a huge body of literature that studies several opti- that matches n users to shared cabs (each with capacity ζ) to
mization problems related to ride sharing [Furuhata et al., maximize profit, given a multiplicative detour-based discount
2013; Agatz et al., 2012; Pelzer et al., 2015], but the prob- fp (δi , τi ), which is a linear function of the distance-wise (δi )
lem of pricing in ridesharing has been relatively less studied, and time-wise (τi ) fractional detours experienced by user i.
with profit optimization rarely discussed. [Banerjee et al., cb is the “base” (fixed) cost for a ride, cd and ct are the costs
2015] characterizes the pros/cons of static versus dynamic per unit distance and time, respectively. fd is the fraction of
pricing. [Jung et al., 2013] provides heuristics for maximiz- driver earnings taken by the service provider.
ing profit subject to bounded detours. In contrast, we study
profit maximization over a pricing model that incorporates a 3.1 ILP Formulation
detour-based discount.
[Nguyen, 2013] proposes a dynamic, demand-responsive Each user i is associated with a source-destination pair
ridesharing system using auction mechanisms, where cus- (Si , Di ). We consider n initially empty cabs and construct
tomers enjoy additive detour-based discounts. Our model a complete directed graph G = (V, E) with 2n vertices
considers multiplicative detour-based discounts; moreover, that represent all the sources and destinations, that is, V =
we also provide an algorithm to learn the profit-maximizing {Si , Di }ni=1 . Each directed edge e = (u, v) represents the
discount parameter under a suitable user choice model. best route from location u to v, and weights wd (e) and wt (e)
A closely related problem is that of cost sharing in peer-to- denote the corresponding distance and time, respectively.
peer ridesharing or carpooling, which, until recently, received The ILP seeks values to assign to a set of optimization
little attention. Individual passengers are either asked to post variables in order to maximize an objective function subject
what they are willing to pay in advance [Cao et al., 2015], to a set of constraints.
to share the total cost proportionately among themselves ac-
cording to the distances travelled [Geisberger et al., 2010; Optimization Variables: Let xi,j,e ∈ {0, 1} denote whether
Agatz et al., 2011], or negotiate their shares on their own or not user i is served by cab j along edge e. Let yj,e ∈ {0, 1}
during/after the ride. Such methods ignore the real-time denote whether cab j travels along edge e while serving one
costs and delays incurred during the ride and are gener- or more users. Let zj ∈ {0, 1} denote whether cab j serves
ally insensitive to the disproportionate delays encountered at least one user. We set xi,j,(u,v) = 0 if (u, v) ∈
/ E.
during the ride. Fair cost sharing in ridesharing has been
studied [Bistaffa et al., 2015; Kleiner et al., 2011; Zhao et Objective Function:
al., 2014; Kamar and Horvitz, 2009; Gopalakrishnan et al.,
2016], but its impact on profitability is left unexplored. We For each user i, we have:
fill this gap by studying the trade-off between average per- (a) Distance travelled, di = n
P P
j=1 e∈E xi,j,e wd (e).
ride profits and market share. Pn P
(b) Time in travel, ti = j=1 e∈E xi,j,e wt (e).
3 Matching Rides to Maximize Profit (c) Distance-wise fractional detour, δi = di −wd (Si ,Di )
wd (Si ,Di )
.
Ridesharing generates significant profit to commercial plat- ti −wt (Si ,Di )
(d) Time-wise fractional detour, τi = wt (Si ,Di )
.
forms such as Uber and Lyft. The profit from a single cab is
the difference between the fare paid by the passengers and the (e) Fare, Ci=(1−fp (δi ,τi ))(cb + cd wd (Si ,Di )+ct wt (Si ,Di )).
For each cab j, we have: 3.2 Heuristic for Real-Time Ride Matching
P
(a) Distance traveled, σj = e∈E yj,e wd (e). Given a set S of users matched to a cab, the driving distance
(b) Time in travel, ηj =
P and time for the best route for serving all the users i ∈ S are
e∈E yj,e wt (e).
denoted by d(S) and t(S), respectively. For each user i ∈ S,
(c) Driver earnings, Ej = (1 − fd )(cb zj + cd σj + ct ηj ).
di (S) and ti (S) denote the driving distance and time from Si
Thus, the objectiveP function to
Pbe maximized is the total to Di along that route, respectively. Thus, the profit from a
profit, given by p = ni=1 Ci − nj=1 Ej . cab serving a set of users S can be computed as follows:
• User Cost: The cost of the ride to user i ∈ S,
Constraints:
Ci (S)=(1−fp (δi (S), τi (S))) (cb +cd di ({i})+ct ti ({i})) ,
(i) yj,e = 1 if and only if xi,j,e = 1 for some i: (1)
yj,e ≥ xi,j,e ∀i, j ∈ {1, 2, . . . , n} ∀e ∈ E where δi (S)= di (S)−d i ({i})
di ({i}) , and τi (S)= ti (S)−t i ({i})
ti ({i}) .
n
• Driver Earnings: The earnings to the driver from the ride,
X
yj,e ≤ xi,j,e ∀j ∈ {1, 2, . . . , n} ∀e ∈ E
i=1 E(S) = (1 − fd ) · (cb + cd d(S) + ct t(S)) . (2)
(ii) zj = 1 if and only if yj,e = 1 for some e: • Profit: The profit to the service provider,
zj ≥ yj,e ∀j ∈ {1, 2, . . . , n} ∀e ∈ E X
X p(S) = Ci (S) − E(S). (3)
zj ≤ yj,e ∀j ∈ {1, 2, . . . , n} i∈S
e∈E
The incremental profit of merging two cabs (combining the
(iii) Each user is picked up by exactly one cab: users matched to two cabs into one and removing the other
n
XX cab from the system), serving sets of users Sj and Sk is:
xi,j,(Si ,u) = 1 ∀i ∈ {1, 2, . . . , n} ∆p(Sj , Sk ) = p(Sj ∪ Sk ) − p(Sj ) − p(Sk ). (4)
u∈V j=1

(iv) Each user is dropped off by exactly one cab: The above framework for real-time ride matching is more
n
XX general than that of the ILP. A limitation of the ILP is that
xi,j,(u,Di ) = 1 ∀i ∈ {1, . . . , n} fp (·) is restricted to be a linear function. Moreover, it may
u∈V j=1
not be possible for an ILP to explicitly incorporate additional
(v) Each user is not served before pickup: detour-aware routing constraints and user preferences. How-
n
XX ever, during performance evaluation of our proposed meth-
xi,j,(u,Si ) = 0 ∀i ∈ {1, . . . , n} ods, we limit the scope to the constraints listed in Section 3.1,
u∈V j=1 in order to compare against the optimal ILP solution.
(vi) Each user is not served after dropoff: Greedy-Max-Profit-Based Method
n
XX We now describe a greedy method for profit-maximizing real-
xi,j,(Di ,u) = 0 ∀i ∈ {1, . . . , n} time matching. We begin by assigning each of n input users
u∈V j=1 to their “own” cab, i.e., no two users are assigned to the same
(vii) Each user is served by a single cab, only between pickup cab.1 We maintain two collections of cabs, namely, a set of
and dropoff: unavailable cabs U and a current pool of cabs A. We initialize
X X A with the n single-user cabs and U = ∅. We repeatedly
xi,j,(u,v) = xi,j,(v,u) ∀v ∈ V \ {Si , Di } ∀i ∀j perform the following steps until A is empty, at which point,
u∈V u∈V
we return U as the output:
(viii) Each cab serves at most ζ users along any edge: (a) For every two cabs j and k in A, if it is feasible to merge
n
X them subject to the capacity constraint (and any other op-
xi,j,e ≤ ζ ∀j ∈ {1, . . . , n} ∀e ∈ E tional constraints), compute the corresponding incremen-
i=1
tal profit pjk = ∆p(Sj , Sk ).
(ix) Each cab’s overall route is not disjoint, i.e., there are (b) Find two cabs j ∗ and k ∗ such that pj ∗ k∗ is maximum. If
no gaps between serving users during which the cab is
empty. Equivalently, if a cab serves a total of k ≥ 1 pj ∗ k∗ < 0, terminate and return U ∪ A as output. Other-
users, then it must do so using exactly 2k − 1 edges: wise, merge them and replace the individual cabs j ∗ and
n X
! k ∗ in A with the merged cab. If the merged cab has ζ
X
yj,e ≥ 2
X
xi,j,(Si ,u) −1 passengers, move it from A to U.
e∈E i=1 u∈V Note that each iteration either terminates the method, or de-
The inequality (≥) is to allow empty cabs (k = 0). creases the size of A by at least one; therefore, the method
terminates after at most n − 1 iterations. The time complex-
Computing the optimal solution to the ILP takes a lot of time
ity of this method is O(n2 log n).
and memory, and is unsuitable for real-time ride matching,
which requires fast, near-optimal techniques. However, it is 1
We assume that all cabs are initially empty; however, our meth-
a handy benchmark against which to experimentally evaluate ods extend to a more general case where some cabs already have
the performance of our proposed methods (Section 3.3). passengers in them with their own constraints to begin with.
(a) Total profit by different methods with small-sized (b) Time taken by different methods with small-sized (c) Total Profit earned using various methods on a large
set of trip requests. set of trip requests. dataset of about 19000 trip requests.

Figure 1: Comparison of various heuristic matching methods according to the total profit and running time.

3.3 Experimental Performance Evaluation ing necessary metrics such as profit, detour, driving distance
In order to compare the performance of our proposed greedy and time. The values cb , cd , ct , and fd are taken from publicly
method, we evaluate them against the optimal solution com- available Lyft user pricing and driver payment for NYC [Lyft-
puted by solving the ILP of Section 3.1. Since the ILP takes Line, 2016a; LyftLine, 2016b].
more than 2 hours on an average (even with only 22 requests) Experimental Results
to output the optimal solution, we first evaluate the near-
We select NYC taxi trips on randomly chosen days of 2013,
optimality of our methods using a small dataset, and then use
between 7:45 - 8:00 pm. We intentionally choose the rush
a larger dataset for demonstrating the scalability and compar-
times, since many shared rides would be possible, and it
ing the speed-optimality trade-off of our proposed heuristic
would be interesting to see how our proposed heuristic per-
and order-based greedy heuristics mentioned in [Biswas et
forms. We run our method on several subsamples by consid-
al., 2017]2 . To compare these order based methods with our
ering a different number of users each time (5, 8, 10, 15, 20,
proposed method, the orderings that we consider are:
and 22). We compare the total profit generated (Figure 1a)
(a) Distance Order: Decreasing order of d(S). and the total time taken (Figure 1b) across each method, av-
(b) Profit Order: Increasing order of p(S). eraged over the small-sized subsamples of users. We observe
that the total profit obtained by the Greedy-Max-Profit-Based
Experimental Setup method is up to 93% optimal and runs 105 times faster than
We evaluate our methods using publicly available New York the optimal algorithm on a machine with a quad Intel Core i7
City (NYC) taxi trip data [Whong and Monroy-Hernández, processor with 32GB RAM. The optimal algorithm becomes
2013]. The 19GB dataset has logs for all NYC trips taken in intractable for instances where the number of users exceeds
2013. Each trip in the dataset specifies pickup and dropoff co- 25 (terminated after over 14.22 hours due to lack of memory).
ordinates, pickup time, dropoff time, trip distance and travel Our proposed algorithm scales to a larger input set
time. Each trip is considered as a “user” for our simulation. with ∼19000 users (obtained from a one-hour slot). The
We assume that each cab serves at most ζ = 3 users. Greedy-Max-Profit-Based method takes ∼35 minutes while
We discretize the underlying NYC map into grids of the Distance-Order-Based and Profit-Order-Based methods
100m2 area and select 2543 representative landmarks using take 7 and 4 minutes, respectively.3 However, figure 1c shows
the methodology of [Rajasubramaniam et al., 2017]. We pre- that the performance of the former is significantly better than
compute the inter-landmark distance and time using Open that of the latter methods in terms of the total profit. Thus, the
Trip Planner [OTP, 2016]. For each user, we map their speed-optimality trade-off becomes important in choosing the
source and destination coordinates to their nearest represen- method that best suits the needs of a service provider. Finally,
tative landmarks and use the precomputed values for comput- we also observe that the Greedy-Max-Profit-Based method
2 matches ∼19000 users into 6701 cabs (reducing the opera-
For completeness, we now briefly describe the greedy heuristics
which are parameterized by an order protocol (originally proposed in
tional cost significantly), obtaining a profit of ∼$134, 500.
[Biswas et al., 2017]). An ordered list L of the initial n single-rider
cabs is created according to a certain order, and the set of unavailable 4 Learning the Optimal Discount Policy
cabs U = ∅ is initialized. At each iteration, the cab at the top of the
list (say j) is picked and merged with the first cab k down the list In the previous section, we laid out efficient ride-matching
such that pjk > 0, subject to the capacity constraint (and any other heuristics to maximize profit under a fixed detour-based dis-
optional constraints). Then, the individual cabs j and k are removed count for users. In this section, our goal is to learn the opti-
from L and the merged cab is inserted at the appropriate position mal detour-based discount policy, under a fixed ride-matching
in L. If no such cab k exists, then j is moved from L to U. If the
3
merged cab is full to its capacity (that is, it has ζ passengers), it If the matching algorithm is invoked once per minute, in NYC,
is moved from L to U. These methods have a time complexity of there are only ∼300 requests per minute on an average, for which all
O(n log n). the greedy methods finish in less than a second.
algorithm. For simplicity, we assume that the detour-based find near-optimal discount parameter, Θ[h∗ ]. The value of op-
discount is a linear function of the fractional distance-wise timal discount parameter depends on P(·) which is the prob-
detour. Thus, the discount to a user i ∈ S is given by ability of opting rideshare. For simulating the user-behavior,
we need to assume P(·) to be any non-decreasing function
fpθ (δi (S)) = tan(θ)δi (S) + b, (5) of θ. In practice, this probability P(θ) is estimated while the
optimal θ is learnt (step 14 of IDFLA). This corresponds to
where b is a constant denoting the minimum discount given the fraction of times (kθ ) the users have opted for ridesharing
to a user (to incentivize them to opt for ridesharing in the whenever the discount parameter was θ. The P(θ) function
first place), and the discount parameter θ ∈ [0◦ , 90◦ ) governs that we have assumed for our simulation, is shown in Fig-
how steeply the discount increases with the detour. Figure 2 ure 3.
plots (5) for various values of θ, assuming b = 10%.
Let P(θ) denote the probability of a user opting to
rideshare, a nondecreasing function of θ. A larger θ would
lead to a larger population of users who are willing to
rideshare, and potentially more shared rides; however, the
profit from each shared ride would be smaller. Since the ex-
pected total profit depends on the number of users opting to
rideshare, as well as the per-ride profit, it is important to find
a value of θ that provides the optimal balance to maximize the
expected total profit. We propose a method, which we call It- Figure 2: Various discount functions. Figure 3: Acceptance probability.
erative Discount Function Learning Algorithm (IDFLA), that
learns such a θ over a period of time. Taxi trips in NYC between 7:45 - 8:00 pm of the 52
Each day, a value of θ is declared by the service provider. Wednesdays of 2013 were used to compare the performance
In response, the users who opt for ridesharing are then of IDFLA with an “oracle” method (BDF) that knows the
matched throughout the day using one of the heuristics from Best Discount Function apriori. Figure 4 shows the average
Section 3.2. The total profit earned is computed at the end daily profit obtained by declaring the same discount function
of the day. IDFLA (Algorithm 1) learns the expected daily everyday. We observe that the best discount function is ob-
profit earned using different values of θ, and eventually con- tained when θ = 40◦ ; therefore, BDF chooses θ = 40◦ every-
verges to the best θ∗ that maximizes this quantity. This day. Figure 5a shows that the discount parameter selected by
kind of sequential decision making in stochastic environment IDFLA converges to that of BDF in 16 days.
can be categorized as a stochastic multi-armed bandit prob-
lem [Bubeck and Cesa-Bianchi, 2012], where the θ values
are the arms, and the daily profit earned by using a certain θ is
the reward earned by pulling the corresponding arm. IDFLA
uses a popular technique for solving stochastic multi-armed
bandit problems, called UCB1 [Auer et al., 2002].

Algorithm 1 IDFLA
1: Input: A finite number of θ values in an array Θ
2: Ph stores the total profit with Θ[h] Figure 4: Average daily profit obtained by various discount functions over 52 days
3: kh stores the number of days Θ[h] is declared
4: for t ← 1 to |Θ| do
5: Choose Θ[t] on tth day and observe profit earned p;
6: Set Pt ← p; Set kt ← 1; 
p
7: Find h∗ = arg maxh Ph + 2 ln |Θ| ;
8: for t ← |Θ| + 1, |Θ| + 2, . . . do
9: Choose Θ[h∗ ] on tth day and observe profit earned p;
10: Update Ph∗ ← Ph∗ + p;  Update q kh∗ ← kh∗ + 1;
11: Find h∗ = arg maxh Ph∗
kh∗ + 2 ln t
kh for (t + 1)th
day;
(a) Comparing daily profit obtained by (b) Average daily profit of IDFLA con-
IDFLA and BDF. verging to that of BDF.

Figure 5: Comparing IDFLA and BDF

4.1 Performance Evaluation of IDFLA
We evaluate our learning method IDFLA under the Greedy- We also observe (Figure 5b) that the average daily profit,
Max-Profit-Based ride-matching heuristic. We take Θ = that is, the cumulative total profit over t days divided by t,
{10◦ , 20◦ , 30◦ , 40◦ , 50◦ , 60◦ , 70◦ , 80◦ }, and run IDFLA to converges to that of the best discount function BDF. We show
that after 52 days, the difference4 in average daily profit ob- slot on that day. We then split the hour into 60 one-minute in-
tained by IDFLA and BDF is just $9.58, or 2.67%. Thus, stances, and collect the sources and destinations for the users
IDFLA quickly learns the optimal discount parameter that who initiated trips in each instance. The average number of
can help the service provider maximize profit while keeping users per instance is 210.85. While reporting our results, we
users sufficiently happy by providing suitable discounts. take an average across these 60 instances.
The parameters cb , cd , ct , and fd are taken from pub-
5 Impact of Detour-Aware Routing licly available Lyft user pricing and driver payment for
NYC [LyftLine, 2016a; LyftLine, 2016b]. We select “detour-
In conjunction with desirable pricing schemes for ridesharing sensitivities” αi (higher values imply more aversion to de-
users, recent work has proposed detour-aware routing poli- tours) randomly from [$0, $5] per mile.We assume that each
cies that impose upper bounds on the total detour [Jung et al., cab serves at most ζ = 2 users, so that the profit maximiza-
2013] (static) or incremental detours [Gopalakrishnan et al., tion problem can be solved optimally in polynomial time us-
2016] (dynamic). Such detour-aware routing policies work ing Edmond’s algorithm [Edmonds, 1965].5 The driving dis-
together with detour-based discount policies in order to en- tance and time between locations are obtained by querying
hance the quality of the ride experience for ridesharing users. Open Trip Planner [OTP, 2016]. The parameters of the linear
This induces a similar trade-off as the one in Section 4: im- detour-based discount policy (5) are θ = 40◦ and b = 10%.
posing additional quality-enhancing constraints into the ride- To begin with, before imposing SIR-γ, we assume that the
matching process incentivizes greater adoption of ridesharing service provider’s initial market share is 60%, within which
and increases the market share of the ridesharing population, each user opts for ridesharing with a probability directly pro-
but could result in reduced average per-ride profit. In this sec- portional to θ and b, and inversely proportional to αi .
tion, we consider the impact of one such detour-aware routing For each instance, we keep the following fixed: user ori-
policy, based on the concept of Sequential Individual Ratio- gins, destinations, whether they are part of the initial market
nality (SIR), introduced by [Gopalakrishnan et al., 2016]. share, and if so, whether they opted for ridesharing. Then, we
SIR guarantees that the disutility to existing users (sum of generate 30 realizations, where each user reacts to the adop-
the monetary cost and an “inconvenience cost” due to de- tion of a detour-aware routing policy (SIR-γ), by flipping bi-
tours) in a shared ride is non-increasing as additional users ased coins parameterized by cin , which is a measure of how
are picked up. A parameterized version of SIR, called SIR- strongly users value a quality-enhancing detour-aware rout-
γ, ensures that the incremental benefit (decrease in disutility) ing policy. This participatory behavior is one of many that
to an existing user upon picking up a new user is at least γ, can be used. For any behavior model, our primary interest is
where γ ≥ 0. Imposing SIR-γ results in a reduced set of the fraction of users who end up opting for ridesharing, and
feasible matched rides, leading not only to reduced average not the generative process itself. In the coin-flip-based pro-
per-ride profit, but also a potential loss due to the “up-front cess, the first group of users, who are in the initial market
discounts” (the parameter b in (5)) given to the passengers share but did not opt for ridesharing, flip with bias pin that
who opt for ridesharing, but end up being unmatched. How- is directly proportional to θ, b, and γ, and inversely propor-
ever, the resulting detour-aware routing policy improves the tional to αi , with cin denoting the constant of proportionality.
quality of shared rides, thereby increasing the market share The second group of users, who are outside the initial mar-
of ridesharing users, which could offset these negative ef- ket share, flip with bias 0.5pin (the factor 0.5 is hand-picked).
fects. Unlike [Gopalakrishnan et al., 2016] that looks at the The users who had already opted for ridesharing remain so.
detour-aware routing policy from the user perspective (espe- For consistency, we ensure that as we increase the value of γ,
cially fairness), in this work we answer such a scheme’s im- only users who have previously not opted for ridesharing flip
pact on the service provider’s profit. again.
We investigate this phenomenon by performing experi- We calculate the profit obtained by matching users in the
ments to address the following questions: new market share. The parameter cin is varied from 100 and
1000, while γ is varied from 0 to 0.9.
(a) What is the trade-off between fewer matched rides and
increased market share, compared to a scenario where no 5.2 Experimental Results
SIR-γ is imposed? How does this affect the profit?
Figure 6 shows profitability as γ increases for various values
(b) How sensitive is the profit to the fraction of additional of cin . The baseline profit, without SIR-γ, is shown as a hor-
users who opt for ridesharing in response to the adoption izontal line. For any value of cin ≥ 200, there is a range of γ
of detour-aware routing by imposing SIR-γ? for which a higher profit can be obtained by imposing the ap-
propriate SIR-γ policy. Moreover, when γ is high, the profit
5.1 Experimental Setup declines, because too many users remain unmatched.
As before, we use the NYC taxi trip data for our experiments. Figure 7 shows a different visualization by plotting prof-
We randomly select a weekday of 2013 and a one hour time itability as a function of cin , for different values of γ.
Figure 8 plots the minimum increase in market share that
4
The difference between the average reward of a learning algo- is necessary to meet a given lower bound on the profit, as
rithm to that of the optimal is called regret. An upper bound on the
5
regret of UCB1 after T iterations is O(log T ) [Auer et al., 2002], The experiments can be carried out for ζ > 2 as well, by using
when each arm’s reward is a bounded random variable. one of our proposed heuristics from Section 3.2 instead.
 Figure 9: Profit versus marketshare increase.
Figure 6: Profit versus γ.

(a) real-time ride matching, (b) user discounts, and (c) detour-
aware routing, by using NYC taxi trip data.
An optimal ride matching algorithm, even on a small set
of 22 requests, can take up to 2 hours, however, our greedy
heuristic generates near-optimal profit in less than 1/105 of
the time, and scales well — it is able to match 18794 trip re-
quests into 6701 cabs — while offering different trade-offs
between optimality and running time. Our adaptive learning
algorithm called IDFLA learns the optimal detour-based dis-
count parameter that achieves the right balance between the
number of passengers who opt for ridesharing and the portion
Figure 7: Sensitivity of profit to cin . of savings taken as profit from each shared ride, within 16 it-
erations. Finally, we demonstrate that even a small market
share increase of 7%, in response to other restricted detour-
a function of γ. For instance, to be 5% more profitable, a aware policies such as sequential individual rationality, is suf-
13% increase is needed, as shown by the green curve. Al- ficient to counter the reduced per-ride profit due to the re-
ternatively, Figure 9 shows the maximum possible increase stricted constraints. Thus, one of our contributions is also a
in profit (over all γ), as a function of the increase in market holistic treatment of the framework for detour-aware policies.
share. The key takeaway here is that a 7% increase6 in market Given the encouraging empirical results of our proposed
share is sufficient to recover from the negative effects (on the solutions for profit optimization, future work should pro-
profit) of adopting a detour-aware routing policy. vide approximation guarantees for the proposed greedy ride-
matching heuristics, and a formal mathematical analysis to
characterize the dependence between profit under a detour-
aware routing policy. Another interesting direction to explore
is the interplay between these different elements, e.g., how
does a detour-aware routing policy affect the performance of
the ride-matching heuristics and the convergence of the learn-
ing algorithm?

References
[Agatz et al., 2011] Niels AH Agatz, Alan L Erera, Mar-
tin WP Savelsbergh, and Xing Wang. Dynamic ride-
Figure 8: Marketshare versus γ for a given profit lower bound. sharing: A simulation study in metro atlanta. Transporta-
tion Research Part B: Methodological, 45(9):1450–1464,
2011.
6 Concluding Remarks [Agatz et al., 2012] Niels Agatz, Alan Erera, Martin Savels-
In this paper, we adopt a profit-centric view of real-time bergh, and Xing Wang. Optimization for dynamic ride-
ridesharing system design, and undertake an empirical in- sharing: A review. European Journal of Operational Re-
vestigation of three key elements that impact profit, namely search, 223(2):295–303, 2012.
6
[Auer et al., 2002] Peter Auer, Nicolo Cesa-Bianchi, and
A 7% increase in market share in response to a quality- Paul Fischer. Finite-time analysis of the multiarmed bandit
enhancing policy is not unrealistic. Lyft’s market share increased problem. Machine Learning, 47(2-3):235–256, 2002.
by over 70% in 2016, owing in large part to better quality of ser-
vice, resulting in Lyft’s customers being happier than those of their [Banerjee et al., 2015] Siddhartha Banerjee, Ramesh Johari,
primary competitor, Uber [Swan, 2016]. and Carlos Riquelme. Pricing in ride-sharing platforms:
 A queueing-theoretic approach. In Proceedings of the Six- taxi dispatch algorithms. In Proc. Transportation Research
teenth ACM Conference on Economics and Computation, Board 92nd Annual Meeting, 2013.
pages 639–639, 2015.
[Kamar and Horvitz, 2009] Ece Kamar and Eric Horvitz.
[Bistaffa et al., 2015] Filippo Bistaffa, Alessandro Farinelli, Collaboration and shared plans in the open world: Stud-
Georgios Chalkiadakis, and Sarvapali Ramchurn. Rec- ies of ridesharing. In IJCAI, pages 187–194, 2009.
ommending fair payments for large-scale social rideshar- [Kleiner et al., 2011] Alexander Kleiner, Bernhard Nebel,
ing. In Proceedings of the 9th ACM Recommender Systems and Vittorio Amos Ziparo. A mechanism for dynamic ride
Conference (RecSys), pages 139–146, 2015. sharing based on parallel auctions. In Proceedings of the
[Biswas et al., 2017] Arpita Biswas, Ragavendran Gopalakr- Twenty-Second International Joint Conference on Artifi-
ishnan, Theja Tulabandhula, Koyel Mukherjee, Asmita cial Intelligence (IJCAI), pages 266–272, 2011.
Metrewar, and Raja Subramaniam Thangaraj. Profit op-
timization in commercial ridesharing. In Proceedings of [LyftLine, 2016a] LyftLine. Lyft Line Driver Pay. https:
the 16th Conference on Autonomous Agents and MultiA- //help.lyft.com/hc/en-us/articles/
gent Systems, pages 1481–1483. International Foundation 213582268-Lyft-Line-Driver-Pay, 2016.
for Autonomous Agents and Multiagent Systems, 2017. [LyftLine, 2016b] LyftLine. Lyft Line Pricing. https:
[Bubeck and Cesa-Bianchi, 2012] Sébastien Bubeck and Ni- //help.lyft.com/hc/en-us/articles/
colo Cesa-Bianchi. Regret analysis of stochastic and 213815178-Lyft-Line-Pricing, 2016.
nonstochastic multi-armed bandit problems, 2012. arXiv
[Nguyen, 2013] Duc Thien Nguyen. Fair cost sharing auc-
preprint arXiv:1204.5721.
tion mechanisms in last mile ridesharing. PhD thesis, Sin-
[Cao et al., 2015] Bin Cao, Louai Alarabi, Mohamed F gapore Management University, 2013.
Mokbel, and Anas Basalamah. SHAREK: A scalable dy-
namic ride sharing system. In Proceedings of the 16th [OTP, 2016] Open Trip Planner. https://
IEEE International Conference on Mobile Data Manage- www.opentripplanner.org, 2016.
ment (MDM), pages 4–13, 2015.
[Pelzer et al., 2015] Dominik Pelzer, Jiajian Xiao, Daniel
[Cordeau et al., 2006] Jean-François Cordeau, Gilbert La- Zehe, Michael H Lees, Alois C Knoll, and Heiko Aydt.
porte, Martin WP Savelsbergh, and Daniele Vigo. Vehicle A partition-based match making algorithm for dynamic
routing. In Handbooks in operations research and man- ridesharing. IEEE Transactions on Intelligent Transporta-
agement science: Transportation, volume 14, chapter 6, tion Systems, 16(5):2587–2598, 2015.
pages 367–428. Elsevier, 2006.
[Rajasubramaniam et al., 2017] T. Rajasubramaniam,
[Edmonds, 1965] Jack Edmonds. Paths, trees, and flowers.
K. Mukherjee, G. Raravi, A. Metrewar, N. Annamaneni,
Canadian Journal of mathematics, 17(3):449–467, 1965.
and K. Chattopadhyay. Xhare-a-ride: A search opti-
[Furuhata et al., 2013] Masabumi Furuhata, Maged mized dynamic ride sharing system with approximation
Dessouky, Fernando Ordóñez, Marc-Etienne Brunet, guarantee. In IEEE International Conference on Data
Xiaoqing Wang, and Sven Koenig. Ridesharing: The Engineering, 2017.
state-of-the-art and future directions. Transportation
Research Part B: Methodological, 57:28–46, 2013. [Swan, 2016] Andy Swan. Sorry, Uber. Social Data Vali-
dates The Lyft Growth Story And Valuation Demands.
[Geisberger et al., 2010] Robert Geisberger, Dennis Luxen,
http://www.forbes.com/sites/andyswan/
Sabine Neubauer, Peter Sanders, and Lars Volker. Fast 2016/08/31/sorry-uber-social-data-
detour computation for ride sharing. In Proceedings of validates-the-lyft-growth-story-and-
the 10th Workshop on Algorithmic Approaches for Trans- valuation-demands/, 2016.
portation Modelling, Optimization, and Systems, OASIcs-
OpenAccess Series in Informatics, volume 14, pages 88– [Whong and Monroy-Hernández, 2013] Chris Whong and
99. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, Andrés Monroy-Hernández. New York City taxi trips.
2010. http://www.andresmh.com/nyctaxitrips,
[Gopalakrishnan et al., 2016] Ragavendran Gopalakrishnan, 2013. Last accessed: 2016-06-18.
Koyel Mukherjee, and Theja Tulabandhula. The costs and [Zhao et al., 2014] Dengji Zhao, Dongmo Zhang, Enrico H
benefits of ridesharing: Sequential individual rationality Gerding, Yuko Sakurai, and Makoto Yokoo. Incentives
and sequential fairness. CoRR, abs/1607.07306, 2016. in ridesharing with deficit control. In Proceedings of the
[Jung et al., 2013] Jaeyoung Jung, R Jayakrishnan, and 2014 International Conference on Autonomous Agents and
Ji Young Park. Design and modeling of real-time shared- Multi-Agent Systems, pages 1021–1028, 2014.