ABSTRACT
Title of thesis: OPERATIONAL CHALLENGES
IN DOCKLESS BIKE-SHARES:
THE CASE OF
HYPERLOCAL-IMBALANCE
Shivam Mishra
Master of Science, 2021
Thesis Directed by: Dr. Ashish Kabra
Department of Decision, Operations
& Information Technologies
and
Prof. Derek Paley
Department of Aerospace Engineering
Recent times have seen a shift from traditional docked to dockless bike-sharing
systems. It is popular among consumers as it allows flexibility to drop-off bikes any-
where and solves the last-mile problem of transportation. While convenient for
users, the dockless bike-share system?s free-floating model introduces the problem
of hyperlocal imbalance, about which little or no research is available. The hy-
perlocal imbalance is the supply-demand disparity created in a small geographical
region due to consumer?s bias to pick up bikes from some locations compared to
others. This paper introduces, demonstrates, and determines the reasons behind
the hyperlocal-imbalance in dockless-bike-sharing systems.
The study of hyperlocal-imbalance requires access to fine-grained trip level data,
which is not easily accessible to the research community due to privacy or compe-
tition issues. To deal with it, in this work, we introduce an algorithm to extract
trip-level information from the General Bikeshare Feed Specification (GBFS) feeds,
which bike-share companies are obliged to upload as per transportation department
regulations across the US. The algorithm is validated against the actual trip data
of dockless bikes. It extracts the trip details from the GBFS data with a recall of
77% and precision of 80%.
OPERATIONAL CHALLENGES IN DOCKLESS BIKE-SHARES:
THE CASE OF HYPERLOCAL-IMBALANCE
by
Shivam Mishra
Thesis submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Master of Science
2021
Advisory Committee:
Professor Derek Paley, Chair/Co-Advisor
Dr. Ashish Kabra, Advisor
Dr. Ilya Ryzhov
? Copyright by
Shivam Mishra
2021
Acknowledgments
I want to thank my advisor, Dr. Ashish Kabra, for providing me with an
opportunity to work on this project. He kickstarted this project last year and al-
lowed me to build on it. Apart from making himself available for help and advice,
he also helped me with all the resources I needed for this project: the workstation,
the books, or anything else. I would also like to acknowledge the financial support
provided by Dr. Kabra while working on this project. It has been a pleasure to
work and learn from Dr. Kabra.
I would also like to thank my co-advisor, Dr. Derek Paley. His advice and sugges-
tions helped me greatly to improve and fine-tune this work. I would also like to
thank him for being the chair of my thesis committee. I also owe my gratitude to
Dr. Ilya Rhyzhov for agreeing to serve on my thesis committee and for sparing the
invaluable time reviewing this work.
ii
Table of Contents
Table of Contents iii
List of Tables v
List of Figures vi
Chapter 1: Introduction 1
1.1 Dockless Bike-Sharing Systems . . . . . . . . . . . . . . . . . . . . . 1
1.2 Supply-Demand Imbalance in Dockless Bikesharing Systems . . . . . 2
1.3 Trip Data Scarcity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Contribution of the Paper . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Chapter 2: Background 6
2.1 General Bikeshare Feed Specification (GBFS) Data . . . . . . . . . . 6
2.2 Hyperlocal Imbalance . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Chapter 3: Extracting trip information from GBFS Data 10
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Algorithm to extract trip info from GBFS data . . . . . . . . . . . . 11
3.4 Comparison with Actual trips . . . . . . . . . . . . . . . . . . . . . . 13
Chapter 4: Hyperlocal Imbalance in Dockless Bike-sharing Systems 19
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Data Collection and Processing . . . . . . . . . . . . . . . . . . . . . 19
4.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2.2 Attributing Pick-ups and Drop-offs to Traffic Intersections . . 20
4.2.3 Cluster of Intersections and Metrics of Interest . . . . . . . . . 22
4.3 Demonstrating Hyperlocal Imbalance . . . . . . . . . . . . . . . . . . 24
4.3.1 Measuring Hyperlocal Imbalance in Washington DC . . . . . . 24
4.3.2 Hyperlocally Imbalanced Clusters vs. Normal Clusters of In-
tersections. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.4 Primary Drivers of Hyperlocal Imbalance . . . . . . . . . . . . . . . . 31
4.4.1 Bike Availability vs. Pick-ups . . . . . . . . . . . . . . . . . . 31
4.4.2 Auxiliary and Central Intersections . . . . . . . . . . . . . . . 33
iii
Chapter 5: Conclusion 37
Bibliography 41
Bibliography 41
iv
List of Tables
2.1 Relative rate and distance from reference intersection values at in-
tersections in the region. Relative rate is the ratio of pickup rate
at any intersection to the pickup rate at the reference intersection
(intersection 1930). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 Columns extracted from GBFS data. . . . . . . . . . . . . . . . . . . 12
3.2 Bike providers and GBFS vehicle ID type. . . . . . . . . . . . . . . . 12
3.3 Metrics measuring performance of Algorithm 2 . . . . . . . . . . . . . 17
4.1 Field of interest in trip and GBFS dataset . . . . . . . . . . . . . . . 20
4.2 Snapshot of information collected for each intersection in Washington
DC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.3 Percentage of clusters/pickups in the given HHI ratio range. . . . . . 27
4.4 Percentage of clusters in the given HHI ratio range for different parts
of day . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.5 Comparison between Central and Auxiliary Intersection of clusters . . 30
4.6 Regression Results with pickups as dependent variable and mean
availability as independent variable . . . . . . . . . . . . . . . . . . . 33
4.7 Snapshot of data used for regression in equation 11 . . . . . . . . . . 35
4.8 Regression Results with TAP(t) as dependent variable and CA(t),
TAA(t) as indepedent variable. . . . . . . . . . . . . . . . . . . . . . 35
v
List of Figures
2.1 Pickup and dropoff percentages at intersections within 270 yards of
the intersection of Constitution Avenue NW and 7th Street NW of
Washington DC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 Trips vs day of the week. . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Trips vs hour of the day. . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Trip duration distribution. . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 Cumulative pickups vs number of intersections . . . . . . . . . . . . . 22
4.2 pdhhi distribution among the clusters in Washington DC. . . . . . . . 27
4.3 cluster normalized pickup rate for intersections among the clusters
with hhi greater and less than 1.1. . . . . . . . . . . . . . . . . . . . . 29
4.4 Mean availability distribution for intersections among the hyperlocally-
imbalanced and normal clusters. . . . . . . . . . . . . . . . . . . . . . 32
vi
Chapter 1: Introduction
1.1 Dockless Bike-Sharing Systems
Almost every major city worldwide has implemented a bike-share system to
make its transportation system smarter and sustainable. In the United States alone,
nearly 200 cities have bike-sharing systems, multiple of them in many cases [1].
Over 136 million trips were completed on shared bikes and scooters in 2019 [2]. The
bike-sharing systems can be divided into docked and dockless categories based on
whether the bikes are picked up from and dropped off at a fixed station or not. Re-
cent years have seen a shift from traditional docked bike-sharing systems to hybrid
or completely dockless systems. The dockless system, in 2019, accounted for 70%
of all shared micro-mobility trips in the US [2]. This is primarily because the dock-
less systems solve the last-mile problem of local transportation. Other than that,
dockless systems improve user experiences, offer high flexibility to the riders, and
provide tighter integration with the public transit system [3]. The ability to drop
off bikes anywhere provides dockless systems an inherent advantage over station-
based systems. This flexibility, however, introduces some significant problems. The
dockless shared bikes are often parked or abandoned in pathways, rivers, and other
public spaces that destroy the city?s aesthetics and create safety hazards and public
1
nuisance. Because of this reason, the number of bikes in the fleet of any dockless
bike-sharing company is regulated by the city?s transportation authority. It is re-
vised regularly, consistent with the city?s needs and performance of the bike-sharing
companies. With a fixed number of bikes in hand, each bike in the fleet needs to
be appropriately managed and optimally utilized to meet the city?s dynamic bike
supply-demands. Bike-providers have to regularly re-distribute the bikes across the
city to keep up with the supply-demand imbalance.
1.2 Supply-Demand Imbalance in Dockless Bikesharing Systems
The most critical aspect of meeting the dynamic supply-demands in the city
is rebalancing. Rebalancing is the distribution of bikes from areas with a surplus
to areas with a shortage of bikes. Rebalancing is critical to meet fluctuations in
bike demands over the city and nearly accounts for 30-50% of the operating costs
for the bike companies [4]. Rebalancing is more difficult in dockless systems where
individual bikes are scattered in an unplanned manner and are often challenging
to locate. The number of bikes at any location for these dockless systems depends
heavily upon user behavior. Their spatial distribution in the city changes after ev-
ery trip and results in a significant supply-demand imbalance. This supply-demand
imbalance can be divided into local and hyperlocal imbalances. The local imbalance
is due to similar bike usage patterns of the users. For example, users mostly ride
from their houses to subway stations during morning peak hours, which leads to
the accumulation of bikes at the subway station and leaves very few bikes in the
2
residential areas. This type of imbalance is generally countered by manual redistri-
bution of the bikes, mostly overnight and sometimes during the day hours. On the
other hand, the hyperlocal imbalance is concentrated in a small geographical area
(few blocks) where users show higher dispersion in the choice of dropoff locations
but a concentrated choice of pickup locations. This phenomenon occurs because
users prefer to pick up from the locations where they are certain of having bikes
available. When dropping off bikes, no such concern arises. This results in bikes
being scattered in a block or two away from favored pickup locations. The likelihood
of bikes being picked up at the other locations is comparatively much lower than the
favoured locations. This hyperlocal imbalance phenomenon is a feature of dockless
bike-sharing systems. The stations in docked systems, which are often more than a
few blocks apart, essentially aggregate pickups and dropoffs and alleviate or elimi-
nate this issue. There is little or no literature available on the hyper-local imbalance
in dockless bike-sharing systems. In this work, we introduce the concept of hyper-
local imbalance, demonstrate this phenomenon using the trip and GBFS (General
Bikeshare Feed Specification) data of dockless bike-sharing companies operating in
Washington DC, and explore the primary reasons behind it.
1.3 Trip Data Scarcity
Conducting research on the user-behavior of bike-sharing consumers requires
access to trip-level data, which bike-share companies are wary of providing due to
competition or privacy issues. Trip-level data is also required by the government
3
authorities to oversee and regulate the bike-sharing operations. Easy access of bike-
sharing trip data to researchers can play a pivotal role in solving complex bike-
sharing problems like rebalancing. In this work, we introduce an algorithm to extract
the trip level information from General Bikeshare Feed Specification (GBFS) data.
GBFS is a specification for real-time, read-only data targeted to provide transit
information to the shared mobility end user [5]. It mainly provides locations of all
parked vehicles at any time and is a must for the bike-providers to make publicly
available for authorities oversight. We validate the trips extracted by the algorithm
in a time period against the actual trips that took place during the same time period.
The data for the actual trips for the brief time period is provided by the DDOTS.
We also use this algorithm to extract trips and study hyper-local imbalance among
the dockless bike-sharing companies operating in Washington DC.
1.4 Contribution of the Paper
This paper makes the following contributions:
1. Propose and validate an algorithm to extract trip level data of bikesharing
systems from GBFS feeds.
2. Introduce the concept of hyperlocal imbalance, demonstrate it using the trip-
data of dockless bikeshare systems operating in Washington DC, and explore
the primary reasons behind it.
4
1.5 Thesis Outline
The thesis is divided into five chapters. Chapter 2 provides a background on
GBFS data and hyperlocal imbalance. Chapter 3 describes the algorithm to extract
trip-level information from the GBFS data feeds. Validation of the algorithm is also
done in this chapter 3. We use the algorithm developed in Chapter 3 to extract trip
data from the GBFS feeds to study the problem of hyperlocal imbalance in chapter
4.
Chapter 4 demonstrates the problem of hyperlocal imbalance in dockless bikes us-
ing the trip data of dockless bikes operating in Washington DC. The trip data is
obtained by extracting trips from the GBFS feed collected for dockless bikes in
Washington DC over a period of one year. We further explore the possible reasons
behind the hyperlocal imbalance in Chapter 4. Chapter 5 concludes the thesis and
discusses the possible future directions.
5
Chapter 2: Background
2.1 General Bikeshare Feed Specification (GBFS) Data
GBFS (General Bikeshare Feed Data) is a specification for real-time, read-only
data targeted to provide transit information to the shared mobility end-user [5]. It
requires the bike-sharing companies to publish data feeds that provide real-time
information on all currently parked bikes? absolute locations. The data feeds need
to be published as JSON files updated at every TTL (Time-to-Live) seconds. The
TTL values generally vary from 30 to 300 seconds [5]. Around 290 bike-share sys-
tems worldwide have adopted the GBFS data standard since its release in November
2015 [5]. In many cities, bike providers must publish the GBFS data feeds to enable
government authorities to oversee and regulate bike companies and uses. GBFS
data is also the backbone for many tools that make shared mobility more accessible
to the user. In this work we introduce another utility of the GBFS data feeds. We
introduce an algorithm in Chapter 3 that can be used to extract trip level infor-
mation from the GBFS data feeds. Trip level data contains origin and destination
pair of bike?s trip along with the trip duration. Origin and destination location are
specified by it?s longitude and latitude values.
6
2.2 Hyperlocal Imbalance
Hyperlocal imbalance refers to the supply-demand imbalance in a small geo-
graphical region (an area of a few blocks) within which the users show a remarkable
propensity to pick up bikes only from some preferred locations. It results in bikes
being consumed rapidly from these preferred locations, whereas bikes at other loca-
tions in the region sit idle for an extended period. Bikesharing companies have to
regularly channel bikes at the preferred locations to keep abreast with the demand
levels, even though the bikes are available for consumers? use at a small walking
distance from the favored spot. The primary reason for this bias is the consumer?s
preference to pick up the bikes from the locations of high bike availability. The users
would know which nearby area would consist of more available bikes to choose from
through their experience or the mobile application.
A hyperlocal imbalance example is illustrated in Figure 2.1 and Table 2.1 using the
dockless bike-sharing trip data in Washington DC. It maps the region in a radius of
270 yards around the traffic intersection of Connecticut Avenue NW and 24th Street
NW in Washington DC. The pickups and dropoffs of dockless bikes in the region are
aggregated over the nearest traffic intersection. The figure shows the pickup (right
panel) and dropoff (left panel) share for bikes attributed to the region?s nearest in-
tersection.
7
Intersection ID 1930 (Ref) 309 5669 7812 259 8004 196
Relative Rate 1 0.17 0.02 0.47 0.22 0.27 0.1
Distance from
0 250 110 175 170 87 230
Ref Int (Yards)
Table 2.1: Relative Rate and Distance from Reference Intersection values at inter-
sections in the region. Relative rate is the ratio of pickup rate at any intersection
to the pickup rate at the reference intersection (intersection 1930).
Figure 2.1: Pickup and dropoff percentages at intersections within 270 yards of the
intersection of Constitution Avenue NW and 7th Street NW of Washington DC.
We notice that at the marked intersection in Figure 2.1, the pickups are much
more concentrated than the dropoffs. The average daily pick-ups at the marked
location is 36, whereas the average daily drop-off is just 20. This means that 16
bikes are peddled in daily at the marked location by the bike-providers to meet
the actual-demand. We also observe from Table 2.1 that the pickup rate at all
8
unmarked intersections is less than half of the pickup rate at the marked intersection.
This points out the under-utilization of the bikes at all unmarked intersections in
the region. Bikes at the unmarked intersections have to wait for a considerably
longer time before being picked up compared to the bikes placed at the marked
location. In Chapter 4, we demonstrate hyperlocal-imbalance in the dockless bike-
share systems among the busiest cluster of intersections in Washington DC using
the HHI (Herfindahl-Hirschman Index) [12].
9
Chapter 3: Extracting trip information from GBFS Data
3.1 Introduction
This chapter introduces an algorithm that extracts trip-level-information from
the GBFS feeds over a period of time. Extracting trip-level info allows researchers
easy access to the fine-grained trip data, which can be used to understand user-
behavior, trip-patterns and improve the city?s bike-sharing infrastructure and effi-
ciency. Trip level info is generally not available to the research community due to
privacy issues. We validate the algorithm and measure its performance by compar-
ing its results with the actual trip info (provided by DDOTS) of bike companies
operating in Washington DC.
3.2 Data Collection
This section details the methods to collect the GBFS data required for the
proposed algorithm. We have collected this data for bike-sharing companies oper-
ating in Washington DC in the time period starting from the second last week of
December in 2019 to the last week of March in 2020. Table 3.1 lists fields scraped
for each of the bike companies. The time interval between subsequent scraping for
10
a bike-provider is thirty seconds. The Vehicle IDs uploaded by bike-providers into
the GBFS feed can either be static or dynamic. Static Vehicle ID means that the
unique identifier for the vehicle would remain the same with time. In contrast, dy-
namic vehicle ID means that the vehicle?s unique identifier would change with every
trip or time-step. Table 3.2 provides the list of companies and bike ID types they
upload into the GBFS feed. GBFS datasets with static vehicle IDs can be utilized
to extract trip-level origin-destination pairs. The trip-level info, then, can be used
to determine the trip-distribution in the city. For the GBFS datasets with dynamic
bike IDs, it is not possible to extract trip origin and destination pairs, but it can
still be used to determine the trip-distribution. By trip-distribution, we mean the
total number of pickups and dropoffs from different parts of the city. We save the
scraped GBFS data into the CSV file format. CSV file at each time-step contains a
list of all the vehicles operating in the city (operated by a particular bike-provider)
along with their unique identifier and location. The location for any parked vehicle
is expressed by its longitude (lon) and latitude (lat) values.
3.3 Algorithm to extract trip info from GBFS data
This section describes an algorithm to extract trip origin-destination pairs
from GBFS data feeds that use static vehicle IDs. GBFS feed reflects a bike?s po-
sition only when it is idle. It disappears from the GBFS feed as soon as any rider
rents it for a trip and re-appears in the feed when the trip is over. The location just
before the disappearance from the GBFS-feed is the trip-origin, and the location at
11
Field Description
Vehicle ID Unique Vehicle Identifier
Latitude Latitude of the parked bike location
Longitude Longitude of the parked bike location
Battery Level Battery level of the parked bike
Vehicle Type e-Scooter or e-Bike
Time Stamp Time at which the data is scraped
TTL Time to Live (in seconds)
Table 3.1: Columns extracted from GBFS data.
Bike Provider Helbiz Jump Razor Skip
Vehicle ID type Static Static Static Static
Bike Provider Spin Lime Lyft Bird
Vehicle ID type Static Dynamic Dynamic Dynamic
Table 3.2: Bike providers and GBFS vehicle ID type.
which it re-appears in the feed is the trip-destination. The time period between the
subsequent disappearing and reappearing of the bike from the GBFS feed is the trip
duration.
To extract trips for a unique bike, we check its location at each time-step available
in the record. If the location of a vehicle changes at (j+ 1)th time-step and distance
between its initial and final location is greater than the threshold distance, we record
the jth and (j + 1)th timestamp as the start and end time of the trip. Similarly, we
record the location of vehicle at jth and j + 1th timestamp as the trip origin and
destination. The algorithm to extract trips from the GBFS feed data is described
12
in Algorithm 2. The helper function, which calculates the distance between two
longitude and latitude pairs, is described in Algorithm 1. The distance calculation
in Algorithm 1 is based on the spherical law of cosines [6].
Algorithm 1: Function to calculate distance between lon and lat pair.
1: procedure distance(lon1, lat1, lon2, lat2)
2: R ? Radius of Earth (6317 km).
3: VS ? Sin(lat1) ? Sin(lat2)
4: VC ? Cos(lat1) ? Cos(lat2) ? Cos(lon1 ? lon2)
5: V ? VS + VC
6: distance ? R ?cos?1(V )
7: return distance
8: end procedure
3.4 Comparison with Actual trips
We validate the trips extracted by Algorithm 2 from the GBFS feed by com-
paring them against the actual trip-data provided by DDOTS. The validation is
done for all the trips occurring in the first week of January 2020 by the vehicles
operated by Skip (bike-share-company) in Washington DC. For clarity, we refer to
the dataset with trips extracted from the GBFS feed as the gbfs-trip dataset and the
dataset with the trips provided by DDOTS as the actual-trip dataset. The valida-
tion process includes comparing the origin-destination pairs in the gbfs-trip dataset
with the origin-destination pairs present in the actual-trip dataset.
13
Algorithm 2: Extract trip info from GBFS data with static Bike IDs.
1: Input GBFS data scraped in time-period ?T, (GBFS?T ).
2: ubid ? Unique Bike IDs in GBFS?T .
3: thr dist ? minimum distance that a trip can have (threshold distance).
4: for i = 1, 2, . . . , length(ubid) do
5: GBFSI ? Records with vehicle id = ubidi in GBFS?T .
6: GBFSI ? Sort GBFSI rows by time (ascending).
7: for j = 1, 2, . . . , length(GBFSI)? 1 do
8: if ((lonj 6= lonj+1) and (latj =6 latj+1)) then
9: if distance(lonj, latj, lonj+1, latj+1) > thr dist then
10: Trip Origin ? location in jth row
11: Trip Destination ? location in j + 1th row
12: Trip Start Time ? timestamp in jth row
13: Trip End Time ? timestamp in j + 1th row
14: Ouput Origin-Destination Pair.
15: end if
16: end if
17: end for
18: end for
To easily compare the origin-destination pairs, we attribute the origin or des-
tination location to the nearest (in terms of distance) available traffic-intersection.
Attributing origin/destination to a traffic-intersection means that we assume that
the trip originated or ended at that particular traffic-intersection instead of the
nearby location. We then compare the trips in the gbfs-trip and actual-trip dataset
based on the day and hour the trip started, starting nearest traffic-intersection
and ending nearest traffic-intersection. [7] provides information about all the traffic-
intersections available in Washington DC, including their location (longitude and
latitude values) and a unique identification ID. Distance between the vehicle?s and
14
traffic intersection?s location is calculated based on their longitude and latitude val-
ues using the spherical law of cosines [6]. Note that we do not attribute any origin
and destination location to a traffic-intersection if the distance between the ori-
gin/destination and nearest traffic-intersection location exceeds 270 yards.
The trips in the gbfs-trip dataset may contain some invalid origin-destination pairs
due to an error in reflecting the bike?s position by the on-board GPS module. The
common reason includes proximity to buildings, trees, bridges, etc., which results in
the GPS signals being blocked or reflected [8]. To eliminate these invalid pairs, we
reject all trips that took place in a very short time period (a period of fewer than 2
minutes), or for a very short distance (distance less than 250m), or at an infeasible
speed (speed > 20 mph).
Another reason for the existence of invalid trips in the gbfs-trip dataset is Re-
balancing. It is typically done by collecting vehicles in the wrong place or at a
deficient battery level into a truck and distributing them over the town after charg-
ing. It generally occurs during the night [9] [10] [11]. To counter the possible invalid
trips introduced due to rebalancing, we reject all the origin-destination pairs occur-
ring late-night and early-morning and focus on the trips during the hours when the
riders are most active (8 AM to 9 PM). We refer to all the steps taken to eliminate
invalid trips from the gbfs-trip dataset as the fine-tuning activities.
The actual trip dataset contains 6480 trips (actual-trips) occurring in the first week
of January 2020 between 8 AM to 9 PM completed by the vehicles operated by
skip (bike-share company) in Washington DC. Algorithm 2 extracts 6154 trips after
fine-tuning activities from the GBFS feed for the same time-period. On comparing
15
the trips in gbfs-trip and actual-trip dataset based on the day and hour the trip
started, starting, and ending traffic intersection ID, we found that 4944 trips in the
gbfs-trip dataset matches exactly with the actual-trip dataset (true-gbfs-trips), the
remaining 1210 trips are false-gbfs-trips generated by the Algorithm. We evaluate
the algorithm?s performance by calculating its precision and recall. Precision is de-
fined as the fraction of generated trips (by the algorithm) that matches the actual
trips, while its recall is defined as the fraction of actual trips generated by the al-
gorithm. We measure the accuracy of the algorithm using the F1 value. F1 value
is the harmonic mean of the calculated precision and recall. Equations 3.1 - 3.3 list
the formula to calculate precision, recall, and F1 values for the algorithm.
Trips generated by Alg. 2 that matches the actual trips
Precision = (3.1)
Total trips generated by Alg. 2
Trips generated by Alg. 2 that matches the actual trips
Recall = (3.2)
Total actual trips
2? Precision? Recall
F1 = (3.3)
Precision + Recall
Table 3.3 lists the value of the metrics measuring the algorithm?s performance
against the actual trip-data.
Figures 3.1 and 3.2 shows how trips are distributed over the week and day.
16
Metric Precision Recall F1-value
Value 0.8033 0.763 0.7826
Table 3.3: Metrics measuring performance of Algorithm 2
For the day, it only shows the hours between 8 AM to 9 PM. We notice from these
plots, that the trips in both the datasets are similarly distributed over the week and
day. The same is true for the distribution of trip-duration in the gbfs-trip dataset
and actual-trip dataset.
Figure 3.1: Trips vs day of the week.
17
Figure 3.2: Trips vs hour of the day.
Figure 3.3: Trip duration distribution.
18
Chapter 4: Hyperlocal Imbalance in Dockless Bike-sharing Systems
4.1 Introduction
Hyperlocal imbalance refers to the supply-demand imbalance in a small geo-
graphical region (an area of a few blocks) within which the users show a propensity
to pick up bikes only from some preferred locations. In this chapter, we demonstrate
hyperlocal-imbalance in the dockless bike-share systems among the busiest cluster of
intersections in Washington DC using the HHI (Herfindahl-Hirschman Index) [12].
We further explore the reasons behind the hyperlocal imbalance among the cluster
of intersection in Washington DC.
4.2 Data Collection and Processing
4.2.1 Data
The analysis is based on the GBFS (General Bikeshare Feed Specification) and
trip data collected for the nine dockless bike-sharing companies operating in Wash-
ington DC in 2020. GBFS publishes the position of all the available bikes/scooters
in real-time. The GBFS feeds are non-archival records, and the data available to us
has been collected over the entire period of the year 2020. The trip data available
19
has been extracted from the GBFS data using the algorithm described in Chapter
3. For this chapter?s purposes, all the analysis is done based on the trip and GBFS
data of one month in 2020, starting from the first week of February to the first week
of March. The fields of interest extracted from the trip and the GBFS data for
processing and analysis in Chapter 4 are listed in Table 4.2.
Trip Data GBFS Data
? Bike ID ? Bike ID
? Trip Start Time ? Time
? Trip End Time ? Bike Location Longitude
? Trip Origin Longitude ? Bike Location Latitude
? Trip Origin Latitude ? Bike Provider (Company)
? Trip Destination Longitude
? Trip Destination Latitude
? Bike Provider (Company)
Table 4.1: Field of interest in trip and GBFS dataset
4.2.2 Attributing Pick-ups and Drop-offs to Traffic Intersections
Although the dockless bikes are scattered in a region, in this work, we at-
tributed the dockless bikes to the nearest traffic-intersection (referred to as inter-
section hereinafter) in Washington DC. Data for the intersections in Washington
DC is obtained from Open Data DC [7], which lists the name, longitude, latitude,
and other attributes for all 8437 intersections in Washington DC. The trip origin
and destination are attributed to the nearest available intersection based on the
20
distance calculated between bike location and intersection using their longitude and
latitude values. Please note that we do not attribute any pickup and dropoffs to an
intersection if the least distance between the pickup/dropoff location and nearest in-
tersection exceeds 270 yards. Attributing pickups/dropoffs to an intersection means
that we assume that the trip originated or ended at that particular intersection in-
stead of the location nearby. This provides us with a method to divide Washington
DC into different sectors, accumulate dropoffs and pickups in these sectors, and
then perform further analysis. The formula used for calculating the distance be-
tween bike and traffic intersection location using their longitude and latitude values
is listed in Equation 4.1. The abbreviations lab, lob, lai, and loi in Equation 4.1
represents Bike Location Latitude, Bike Location Longitude, Intersection Latitude,
and Intersection Longitude respectively. We collect the total number of pickups and
dropoffs at each of the 8437 intersections present in Washington DC.
distance = arccos (sin (lab)sin(lai) + cos (lab) cos (lai) cos (lob? loi)) (4.1)
Out of the 8437 intersections in DC, not all intersections are attributed to
a significant number of pickups. Figure 4.1 plots the cumulative pickups vs. the
number of intersections. We notice that out of the total 8437 intersections, ap-
proximately 2000 intersections accounts for almost all of the dockless pickups in
Washington DC. For our purpose, we would focus only on these 2000 intersections.
21
Figure 4.1: Cumulative pickups vs number of intersections
4.2.3 Cluster of Intersections and Metrics of Interest
Many of the intersections in Washington DC, which are very close to each
other (at a distance of almost one or two blocks away), are consolidated into a
cluster. Almost all of these clusters have one popular intersection that accounts
for the majority of the pickups in the cluster; this intersection is referred to as the
reference or the central intersection of the cluster. All the other intersections of the
cluster are referred to as auxiliary intersections. The average distance of auxiliary
intersections from the reference intersection in their corresponding cluster is around
200 yards (note that the average block length is 220 yards).
At each of the intersections in Washington DC, we use GBFS data to calculate the
bikes? mean availability. Please note that the GBFS data provides the real-time
location (longitude and latitude values) of all stationary (bikes that are not engaged
in any trip) dockless bikes. We again attribute the location of each bike to the
22
nearest available intersection. At each time-step, we calculate the total number
of bikes at each intersection. We then use this to calculate the average number
of bikes available at each intersection (mean availability). Using the GBFS data,
we also calculate the time-period for which at least one bike is available at each
intersection (available time). Along with the bike availability and time available, we
also calculate pickup rate, relative rate, and relative availability for each intersection.
Equations 4.2, 4.3, and 4.4 lists the formulas to calculate the pickup rate, relative
rate, and relative availability at intersection k (intk) belonging to the cluster K
(clusterK), which has ref intK as the reference/central intersection.
Total Pickups (intk)
Pick-up Rate (intk) = (4.2)
Available Time (intk)
Pickup Rate (intk)
Relative Rate (intk) = (4.3)
Pickup Rate (ref intK)
Mean Availability (intk)
Relative Availability (intk) = (4.4)
Mean availability (ref intK)
Table 4.2 presents the snapshot of the information accumulated for each of the
intersections in Washington DC using the GBFS and trip data.
23
Int Cluster Mean Pickup Rel. Rel.
Lon Lat Pickups Dropoffs
ID ID Avail Rate Rate Avail
1930 -77.0523 38.9249 C1930 13 674 427 0.03 1 1
196 -77.0533 38.9266 C1930 1 45 48 0.003 0.1 0.08
259 -77.0526 38.9235 C1930 3 135 118 0.007 0.22 0.230
Table 4.2: Snapshot of information collected for each intersection in Washington
DC
4.3 Demonstrating Hyperlocal Imbalance
4.3.1 Measuring Hyperlocal Imbalance in Washington DC
This section focuses on demonstrating the hyper-local imbalance in dockless
bike-sharing systems using the information collated in section 4.2.3 about the pickup
and dropoff numbers at each intersection in Washington DC. We quantify this imbal-
ance using the HHI (Herfindahl-Hirschman Index) metric [12]. HHI is a commonly
accepted metric to calculate the market concentration; it is calculated by squaring
each competing firm?s share in the market and then summing the resulting number.
For our purposes, we calculate the pickup and dropoff HHI by squaring the share of
pickups/dropoffs for each intersection in the cluster and then summing the resulting
numbers. The pickup and dropoff HHI gives information about the concentration
level of pickups and dropoffs in the cluster. At the same time, their ratio quantifies
the contrast between pickup and dropoff distribution among the cluster?s intersec-
tions. Equations 4.5 to 4.7 summarizes the calculation of pickup HHI, dropoff HHI,
and the HHI ratio for a cluster with a total of N intersections. pn and dn are the
24
pickup and dropoff share of the nth intersection in the cluster.
?N
Pickup HHI = p2n (4.5)
n=1
?N
Dropoff HHI = d2n (4.6)
n=1
?
Pickup HHI N p2
HHI Ratio = = ?n=1 n
Dropoff HHI N
(4.7)
2
n=1 dn
Ideally, pickup and dropoff distribution in any cluster should be equal. The
most efficient bike-sharing system would have an equal number of pickups and
dropoffs at each intersection, which means an HHI ratio of 1 at each cluster. How-
ever, in general, this is not the case. The pickup and dropoff distribution vary sub-
stantially. For the hyperlocally-imbalanced cluster, pickups are more concentrated.
We can use the HHI ratio to calculate the percentage difference (pd) by which clus-
ter?s pickups are more concentrated than the cluster?s dropoffs. The formula for the
same is listed in equation 4.8.
pd = (HHI ratio? 1)? 100 (4.8)
Figure 4.2 shows the HHI-ratio distribution among all the clusters in Wash-
ington DC. It shows that the pickups are more concentrated for most clusters than
25
the dropoffs (i.e., HHI ratio > 1). We focus only on the clusters where pickups are
at least 10 percent more concentrated than the dropoffs. We refer to these clusters
as hyperlocally-imbalanced clusters (HIC). Other clusters are referred to as normal
clusters (NC). Around 30% of all clusters falls into the hyperlocally imbalanced
category. Table 4.3 summarizes the distribution of HHI ratios among these clus-
ters. It shows the Percentage of clusters in a given HHI ratio range. On average,
the pickups for these hyper-locally imbalanced clusters are 20% more concentrated
than the dropoffs. Table 4.3 also shows the percentage of all pickups coming out
from these clusters to manifest their importance. It shows that around 36% of all
pickups in Washington DC are from the clusters which suffer from hyperlocal im-
balance. We further explore how the disparity in pickup and dropoff concentration
varies throughout the day. Table 4.4 lists how the HHI-ratio varies through different
parts of the day for the hyperlocally imbalanced clusters. From Table 4.4, it can be
noted that there is not much difference in the distribution of HHI ratios throughout
the day. The total percentage of hyperlocally imbalanced clusters remains almost
the same for each part of the day, highlighting that even with consumer?s different
pickup behavior with the time of day [15], the problem of hyperlocal imbalance
among the cluster of intersections remains the same.
26
Figure 4.2: pdhhi distribution among the clusters in Washington DC.
HHI-ratio [1.1, 1.2) [1.2, 1.3) [1.3 ++] Sum
% of clusters 16.12 8.68 4.13 28.93
% of pickups 21 10 5 36
Table 4.3: Percentage of clusters/pickups in the given HHI ratio range.
4.3.2 Hyperlocally Imbalanced Clusters vs. Normal Clusters of In-
tersections.
As mentioned in Section 4.2.3, the clusters are divided into central and auxil-
iary intersections, where the central intersection carries the most number of pickups
in the cluster. For the hyperlocally imbalanced clusters (HIC), the pickup rate of
the bikes at the majority of the auxiliary intersection is meager as compared to the
central intersection, which points out that the bikes at the auxiliary intersections of
the HIC have to wait for a considerably longer period of time before being picked
27
HHI ratio all day morning late morning afternoon evening late evening
[ 1.1, 1.2 ) 16.12 13.22 19.01 20.25 19.01 16.53
[ 1.2, 1.3 ) 8.68 7.44 9.09 7.85 8.68 10.74
[ 1.3++) 4.13 11.16 6.61 4.96 3.72 2.48
Sum 28.93 31.82 34.71 33.06 31.5 29.75
Table 4.4: Percentage of clusters in the given HHI ratio range for different parts of
day
up as compared to the central intersection of the HIC. The low pickup rate of bikes
at any location leads to low utilization of bikes, which reduces the system?s over-
all efficiency. To exhibit this phenomenon, we use relative rate (see Section 4.2.3).
Relative rate or cluster normalized pickup rate is calculated for any intersection by
dividing the intersection?s pickup rate by the pickup rate of the central intersection
of the cluster to which it belongs. Figure 4.3 shows the distribution of relative rates
of auxiliary intersections for both the normal and hyper-locally imbalanced clusters.
It can be noted that a high percentage of auxiliary intersections in hyperlocally im-
balanced clusters have very low relative pickup rates, unlike normal clusters where
we see a symmetric distribution of relative rates. It reveals the selective behavior of
users in hyperlocally imbalanced clusters who favors the central intersection more
dominantly, unlike consumers in the normal cluster who are open to picking up bikes
from the auxiliary intersection as well.
Table 4.5 summarizes the comparison between the hyperlocally imbalanced
28
Figure 4.3: cluster normalized pickup rate for intersections among the clusters with
hhi greater and less than 1.1.
and normal clusters based on their relative pickup rates at the auxiliary intersec-
tions. Note that around 49% of auxiliary intersections among the hyperlocally im-
balanced clusters have had a relative pickup rate of less than 0.25. In other words,
on average, for each of the hyperlocally imbalanced clusters, almost half of its aux-
iliary intersections have their pickup rate less than a quarter of the pickup rate at
the central intersection, a location which on average is just 200 yards away. We also
note that auxiliary intersections in the hyperlocally imbalanced clusters account for
around 33% of total dropoffs and around 37% of the total availability. These bikes
would have been better utilized if they had been placed just a block away. For hy-
perlocally imbalanced clusters, the bike operators have to regularly channel-in bikes
at the central intersection to meet their demand values even when usable bikes are
available just one block away. This is not the case with the normal clusters, as
their auxiliary intersections with relative pickup rates less than 0.25 account only
29
for 13% of the total dropoffs compared to the mammoth one-third of the dropoffs
for the hyperlocally imbalanced clusters. The users in the hyperlocally imbalanced
clusters drop off the bikes at locations from where the probability of being picked
up is comparatively minimal.
Hyperlocally
Relative
Imbalanced Normal clusters
Pickup Rate
clusters
% of intersections in
x = 0.50 83 % 56 %
cluster where
x = 0.25 49 % 23 %
relative pickup rate
x = 0.10 9.8 % 6.9 %
is less than x.
% of dropoffs at the
x = 0.50 64.7 % 44.4 %
intersections where
x = 0.25 32.9 % 13.3 %
relative pickup rate
x = 0.10 4.6 % 2.8 %
is less than x.
% of bikes available
at the intersections x = 0.50 68 % 45.45 %
in the cluster where x = 0.25 37 % 14.77 %
relative pickup rate x = 0.10 6.5 % 4.1 %
is less than x.
Table 4.5: Comparison between Central and Auxiliary Intersection of clusters
30
4.4 Primary Drivers of Hyperlocal Imbalance
4.4.1 Bike Availability vs. Pick-ups
We plot the relative availability distribution among auxiliary intersections for
both the clusters to analyze why relative pickup rates are very low for auxiliary
intersections in the hyperlocally imbalanced clusters compared to the normal clusters
in Figure 4.4. On average, a greater number of auxiliary intersections with very
low relative mean availability are present in the hyperlocally imbalanced clusters
compared to the normal clusters. Around half of the auxiliary intersections in the
hyperlocally imbalanced cluster has relative availability less than 0.2, compared to
the quarter of the auxiliary intersections in the normal clusters. The primary reason
for this is bikes being more spatially scattered in hyperlocally imbalanced clusters
than in normal clusters. Note that, on average, dockless bikes are aggregated over 5
and 8 intersections in normal and hyperlocally imbalanced clusters. Greater spatial
scattering in hyperlocally imbalanced clusters does not allow the bikes to reach a
threshold bike-availability-level at several intersections necessary for consumers to
feel confident about picking up the bikes. People are generally wary of bikes being
hard to locate, being non-operational, or someone else reaching before them to pick
up the bike first. Hence, they may prefer either to pick up the bikes from some
high-availability intersection or use other modes of transport.
We have now identified that the hyperlocal imbalance and mean availability
are closely associated. Consumers? over-inclination towards high bike availability
31
Figure 4.4: Mean availability distribution for intersections among the hyperlocally-
imbalanced and normal clusters.
sites is one of the major reasons they skip picking bikes from some locations and
favor picking up from other locations. These sites with low mean bikes? availability
are known to the user through the bike company?s mobile application or experience.
The over-inclination prevents the users from picking bikes from intersections in the
cluster with low mean availability. This also forces the bike-sharing companies to
regularly channel bikes to popular locations, even though bikes may be available in
a nearby region (region of one or at max 1.5 blocks away) for pickup. To investigate
this relationship between bike availability and pickup numbers at intersections, we
conducted regression analysis. We tried multiple regression models with pickups
and intersections? mean availability as the dependent and independent variables.
We wanted to find a relationship between them, which can help us deduct how
the pickups at an intersection vary with their mean availability. The model which
stood out in the regression analysis used log-transformed dependent and independent
32
variables. The regression equation is expressed in Equation 4.9.
ln(pickups) = ?? ln(mean availability) + ? (4.9)
No. Observations R-square Adj R-square ? ?
1.4476* 3.2129*
1176 0.466 0.465
(0.045) (0.047)
Table 4.6: Regression Results with pickups as dependent variable and mean avail-
ability as independent variable.
Note : *p < 0.001
The regression analysis is based on the total number of pickups and the mean
availability at intersections in one month. Results for the regression are listed in
Table 4.6. We modify the results from Table 4.6 in equation 4.10 to evaluate daily
pickups. As per equation 4.10, at any intersection, a 10% increase in mean avail-
ability results in a 15% increase in the number of pickups. This disproportionate
increase in pickups with an increase in mean availability is the root of the hyperlocal
imbalance.
24.85
daily pickups = ?mean availability1.4476 (4.10)
30
4.4.2 Auxiliary and Central Intersections
All the analysis we performed using the pickups, dropoffs, relative rates, and
availability indicates that the consumers prefer picking up bikes at the central loca-
33
tion instead of the auxiliary location for all intersections. In this section, we exhibit
this by deriving a relationship between the pickups at the auxiliary intersections
of the cluster with the cluster?s central intersection?s availability values. The rela-
tionship is derived by dividing the data into different time-periods (time-period of
fifteen minutes) and collecting the following fields of interest for each time-period
using the GBFS and trip data.
1. Total Aux. Pickups at time t (TAP(t)):
Sum of pickup values at all the auxiliary intersections of the cluster.
2. Total Aux. Avail. at time t (TAA (t)):
Sum of mean bikes available at all the auxiliary intersections of the cluster.
3. Central Avail. at time t (CA (t)):
Mean availability at the central intersection of the cluster.
4. Cluster ID:
Unique identifier identifying the cluster of the intersection to which each row
of data belongs.
5. Time:
Starting time of the fifteen minutes time period for which the data is collated
in the row.
34
Time TAP(t) TAA(t) CA(t) Cluster ID
2020-02-07 17:45:00 0.0 5.36 2.70 C1930
2020-02-07 18:00:00 1.0 8.15 2.38 C1930
2020-02-07 18:15:00 0.0 8.15 2.10 C1930
Table 4.7: Snapshot of data used for regression in equation 11
TAP(t) = ? CA(t) + ? TAA (t) + ? (4.11)
A snapshot of the data used for the regression is available in Table 4.7. Please
note that we rearrange all the available one-month data to get the above-described
fields of interest for each cluster at 15 minutes span. The regression equation is
listed in Equation 4.11, and the results for the regression are listed in Table 4.8 .
Note that we performed the regression after dropping the duplicate rows and with
the cluster-ID fixed effects.
No. Observations R-square Adj R-square ? ?
-0.0124* 0.0362*
36470 0.201 0.200
(0.002) (0.001)
Table 4.8: Regression Results (Note : *p< 0.001) with TAP(t) as dependent variable
and CA(t), TAA(t) as indepedent variable.
From Table 3.9, we observe that the coefficient for the CA(t) (availability at
35
the center intersection of the cluster) is negative. This shows that the pickups at
the auxiliary intersections decrease with an increase in availability at any cluster?s
central intersection. The negative coefficient for the CA(t) conforms with the obser-
vations and analysis conducted so far in this chapter. It reveals that availability at
its central intersection significantly affects the pickups at its auxiliary intersections
for a cluster.
36
Chapter 5: Conclusion
This thesis proposes and validates a novel algorithm to extract the trip-level
origin-destination pairs from the GBFS dataset. It also introduces the concept of
hyperlocal imbalance in dockless bikeshare, demonstrates its presence using the trip
dataset of dockless bikes operating in Washington DC, and explores the primary
reasons behind it. Chapter 2 provides a background on the GBFS data and hyper-
local imbalance. Chapter 3 introduces an algorithm to extract trip data from the
GBFS feeds with static bike IDs with a great accuracy. Trip level information of
dockless bike-sharing companies is not easily available to the research community
due to privacy issues. Bike-providers often treat it as a commodity rather than a
public resource that could enable users and cities to benefit from city movements and
infrastructure planning [13] [14]. The capacity to extract trip information or trip
distribution from the GBFS feed greatly improves the accessibility of trip-level data;
this enables researchers and government authorities to optimize the bike-sharing sys-
tems? overall efficiency.
We also validated the algorithm by comparing the trips extracted from GBFS data
with the actual available trip data provided by DDOTS. The algorithm was able
to re-generate around 77% of all actual trips (Recall = 77%). Around 20% of the
37
algorithm?s trips did not correspond to any of the actual trips (Precision = 80%).
Possible sources for wrong trip generation are rebalancing of bikes and
error in location reflection by the on-board GPS. We can reduce this number
by using more refined tricks to eliminate invalid trips. One approach could be using
the battery levels before and after a trip to eliminate the invalid trips, but not all
bike-providers upload the battery levels in GBFS feeds. Future work includes refin-
ing the fine-tuning tricks to eliminate all invalid trips generated by the algorithm
(i.e., to achieve 100% precision).Future work also includes a wide variety of analyses
on the generated trip-data; the first is to analyze dockless bike riders? user behavior
in Washington DC. We will use the generated trip data to compare the performance
of dockless bikes against the docked bikes. We also plan to explore the trip patterns
in equity emphasis areas in Washington DC; specifically, we want to explore how
bikes? usability in these areas compare against the rest of the city. We also want
to find out whether the regulations on bike distribution put forward by DDOTS in
Washington DC really help the equity emphasis area?s residents or not.
Chapter 4 introduces the concept of Hyperlocal Imbalance in dockless bikes. It is
defined as the supply-demand imbalance occurring in an area of few blocks due to
the user?s bias to pick up bikes from some preferred locations and comfort to drop
off bikes anywhere they want. There is no research available at present on the hy-
perlocal imbalance among dockless bikes. We demonstrate this phenomenon using
the dockless bike data of bike-sharing companies operating in Washington DC. We
found that almost 30% of all the clusters in Washington DC are suffering from hy-
perlocal imbalance.
38
Although the HHI metric is robust in measuring the hyperlocal imbal-
ance of bikes, this thesis doesn?t provide any insight about it?s statistical
significance. Future work includes analysis to determine the significance
of HHI values measured for cluster of intersections. Another limitation
for the HHI analysis conducted in the thesis comes from the data used
for this purpose. The data available for the HHI analysis was just for one
month (Feb 2020 to March 2020) and hence the calculated HHI metrics
do reveal the state of hyperlocal imbalance during the low temperature
months, when comparatively less number of people use dockless bikes.
A better calculation of HHI metric should include data from all seasons
(preferably the whole year data). Future work includes doing a wider
temporal data collection (that includes all seasons) to calculate more ro-
bust HHI-metrics.
We also quantified the impact of hyperlocal imbalance on bike utilization using the
pickup rate metric. We found that almost one-third of the bikes in the hyperlocally
imbalanced clusters are placed at locations where their utilization is less than a
quarter of what it would have been if they were placed just a block away. We also
displayed that greater scattering of bikes in these clusters prevents bikes? availability
at intersections to reach a level necessary for a user to feel confident about picking
up the bikes. We also derived a regression relationship between pickups and avail-
ability at any intersection which reveals the user?s over-preference towards the sites
with high mean availability to pick up the bikes. A 10% increase in bike availability
at any location results on average increases pickups by 15%.
39
We demonstrated through Washington, DC data that the auxiliary inter-
sections in hyperlocally imbalanced clusters suffer from very low mean
availability of bikes. The low bike-availability-level at these intersections
doesn?t make consumers comfortable enough to pickup bikes from there.
Consumers do not prefer to walk up to these locations to pick up bikes as
they are wary of bikes being damaged, dirty, disabled or being picked up
by someone else before they reach there. Possible ways to alleviate this
issue includes providing a pickup and drop-off incentive; bike-providers
can provide incentive to consumers for picking up and dropping off bikes
at low availability locations. Pickup incentive would enable the bikes at
low availability locations to be better utilized, while dropoff incentive
would allow the bikes at low availability locations to reach the threshold
level required by the consumers to pick up the bikes. Bike-providers can
also allow consumers to reserve the bikes from their mobile application
so that they won?t have to worry about someone else taking that bike
before.Future work includes developing more solutions to curb the issue
of hyperlocal imbalance.
40
Bibliography
[1] Bureau of Transportation Statistics (US Department of Transportation, 2019)
[2] Shared Micromobility in the US (NACTO, 2019)
[3] Zhang, Yuanyuan and Zhang, Yuming Associations between Public Transit Us-
age and Bikesharing Behaviors in The United States (In: Sustainability 10.6
(2018), p. 1868)
[4] The Bikesharing planning guide (ITDP, 2018)
[5] General Bikeshare feed specification (NABSA, 2020, URL
https://github.com/NABSA/gbfs)
[6] Weisstein, Eric W., Spherical Trigonometry (Wolfram Web Resource, URL:
https://mathworld.wolfram.com/SphericalTrigonometry.html)
[7] Intersection Points, Washington DC (Open Data DC, 2019, URL :
https://opendata.dc.gov/datasets/96a9bbbb475648769e311d03c78698a7 2)
[8] How Accurate is GPS (GPS.gov, National Coordination Of-
fice for Space-Based Positioning, Navigation, and Timing, URL:
https://www.gps.gov/systems/gps/performance/accuracy/)
[9] Chemla, D.; Meunier, F.; Calvo, R.W., Bike sharing systems: Solving the static
rebalancing problem (Discrete Optimization 2013, 10, 120?146)
[10] Pal, A.; Zhang, Y., Free-floating bike sharing: Solving real-life large-scale static
rebalancing problems. (Transportation Research Part C: Emerging Technologies
80 (2017): 92-116)
41
[11] Erdo ?gan, G.; Battarra, M.; Calvo, R.W., An exact algorithm for the static
rebalancing problem arising in bicycle sharing systems (European Journal of
Operational Research 245.3 (2015): 667-679)
[12] Rhoades, Stephen A, The herfindahl-hirschman index (Fed. Res. Bull.
79(1993):188)
[13] Go?ssling, S., ICT and transport behavior: A conceptual review (International
Journal of Sustainable Transportation, 12(3), 153?164, 2018)
[14] Spinney, J., Lin, W.-I, Are you being shared? Mobility, data and social relations
in Shanghai?s public bike sharing 2.0 sector (Applied Mobilities, 3(1), 66?83,
2018)
[15] Zhou, Xiaolu, Understanding spatio temporal patterns of biking behavior by an-
alyzing massive bike sharing data in Chicago (PloS one 10.10 (2015): e0137922)
42