Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing

Abstract

Millions of people use platforms such as Google Maps to search for routes to their desired destinations. Recently, researchers and mapping platforms have shown growing interest in optimizing routes for criteria other than travel time, e.g. simplicity, safety, and beauty. However, despite the ubiquity of algorithmic routing and its potential to define how millions of people move around the world, very li"le is known about the externalities that arise when adopting these new optimization criteria, e.g. potential redistribution of traffic to certain neighborhoods and increased route complexity (with its associated risks). In this paper, we undertake the first controlled examination of these externalities, doing so across multiple mapping platforms, alternative optimizations, and cities. We find, for example, that scenic routing (i.e. " beauty "-optimized routing) would remove vehicles from highways, greatly increase traffic around parks, and, in certain cases, do the same for high-income areas. Our results also highlight that the interaction between routing criteria and urban structure is complex and effects vary from city to city, an important consideration for the growing literature on alternative routing strategies. Finally, to address the lack of open implementations of alternative routing algorithms and controlled routing evaluation frameworks, we are releasing our alternative routing and evaluation platform with this paper.

Full text

!Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
Beautiful…but at What Cost? An Examination of Externalities in
Geographic Vehicle Routing
I. JOHNSON, Northwestern University
J. HENDERSON, University of Minnesota
C. PERRY, University of Minnesota
J. SCHÖNING, University of Bremen
B. HECHT, Northwestern University1
Millions of people use platforms such as Google Maps to search for routes to their desired destinations. Recently, researchers
and mapping platforms have shown growing interest in optimizing routes for criteria other than travel time, e.g. simplicity,
safety, and beauty. However, despite the ubiquity of algorithmic routing and its potential to define how millions of people
move around the world, very li"le is known about the externalities that arise when adopting these new optimization criteria,
e.g. potential redistribution of traffic to certain neighborhoods and increased route complexity (with its associated risks). In
this paper, we undertake the first controlled examination of these externalities, doing so across multiple mapping platforms,
alternative optimizations, and cities. We find, for example, that scenic routing (i.e. “beauty”-optimized routing) would
remove vehicles from highways, greatly increase traffic around parks, and, in certain cases, do the same for high-income
areas. Our results also highlight that the interaction between routing criteria and urban structure is complex and effects vary
from city to city, an important consideration for the growing literature on alternative routing strategies. Finally, to address
the lack of open implementations of alternative routing algorithms and controlled routing evaluation frameworks, we are
releasing our alternative routing and evaluation platform with this paper.
CCS Concepts: • Human-centered computing → Ubiquitous and mobile computing;
General Terms: Geography; Algorithms; Vehicle Routing
Additional Key Words and Phrases: externalities; alternative routing; urban structure
ACM Reference format:
Isaac Johnson, Jessica Henderson, Caitlin Perry, Johannes Schöning, and Brent Hecht. 2017. Beautiful…but at What Cost?
An Examination of the Externalities in Geographic Vehicle Routing. PACM Interact. Mob. Wearable Ubiquitous Technol.
1, 2, Article 15 (June 2017), 21 pages.
DOI: 10.1145/3090080
1 INTRODUCTION
The simple act of driving from one place to another is an incredibly common part of many people’s lives. However, it is also
a surprisingly complex task: in many cases, there are seemingly countless routes between a given origin and destination
pair. While the predominant focus of the literature and applications in the geographic routing domain has historically been
on minimizing travel time or distance (e.g. [3,13]), researchers and practitioners have recently shown interest in alternative
routing criteria. For instance, researchers have developed routing systems that generate “scenic” routes (e.g. [41,44,59]),
simpler routes (e.g. [7,47]), and safer routes (e.g. [11,21,45]), among other alternative route optimizations (e.g. [23,25,48,60]).
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
This work is supported in part by the National Science Foundation (CAREER IIS-1707296 and IIS-1707319) and a Lichtenberg
Professorship from the Volkswagen Foundation.
Authors’ addresses: I. Johnson, Northwestern University, 2240 Campus Drive, Frances Searle Building 2-431, Evanston, IL, USA, 60208; B.
Hecht, Northwestern University, 2240 Campus Drive, Frances Searle Building 1-144, Evanston, IL, USA, 60208.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that
copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first
page. Copyrights for third-party components of this work must be honored. For all other uses, contact the Owner/Author.
Copyright 2017 held by Owner/Author
2474-9567/2017/6-ART15
DOI:http://dx.doi.org/10.1145/3090080
15

15:2 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
Similarly, the routing platform Waze has begun to suggest routes, at least in Rio de Janeiro, that avoid areas deemed to have
higher rates of violence [39], and Microsoft owns a patent for pedestrian routing that avoids “unsafe” areas based on crime
and weather factors [51].
Despite this growing interest in alternative routing strategies, there has been no controlled evaluation of the externalities
that arise when more traditional optimization criteria such as travel time are superseded by new optimization criteria such
as safety or beauty. Evaluations of these new criteria have solely considered the direct trade-off with travel time or distance,
e.g. increased travel time to achieve more scenic or safer routes. These evaluations, however, miss the externalities that may
arise with these new criteria, externalities that may have significant social, economic, and safety implications. For instance,
at the community level, these routing approaches may lead to increased or decreased traffic in certain areas. Additionally,
at the route level, these approaches may lead to routes with more turns (directly contradicting user preferences [14,27],
increasing driver stress [56] and cognitive load [16,36], and potentially decreasing safety [33,39]). Moreover, anecdotal
evidence suggests that these externalities may be substantial. Consider, for instance, the widespread outcry about increased
traffic and noise in previously out-of-the-way neighborhoods attributed to routing algorithm changes by Waze (e.g.
[18,29,57]).
In this paper, we aim to address this gap in the literature. To do so, we developed a controlled experimental routing
platform and used this platform to investigate externalities that arise with three common approaches to alternative routing:
scenic routing, safety routing, and simplicity routing. Specifically, examining routes from four cities – San Francisco, New
York City, London, and Manilla – we ask the following research questions:
RQ1: Does optimizing on alternative criteria in routing algorithms lead to route-level externalities such as more complex
routes?
RQ2: Does optimizing on alternative criteria in routing algorithms lead to community-level externalities such as increased
or decreased traffic in certain areas?
Additionally, as noted above, at least one routing platform (Waze) has already implemented alternative routing
techniques [28], and Microsoft has a patent on similar approaches [51]. As such, we also saw an opportunity to use our
experimental platform to better understand (and track) the criteria on which popular routing platforms are optimizing and
to assess whether there are any externalities associated with these criteria. Thus, in the tradition of the algorithmic auditing
literature (e.g. [2,4,53]), we also asked a third research question:
RQ3: Is there evidence of alternative routing criteria being used by popular routing platforms? If so, what are the
associated externalities?
Overall, we find that there are large externalities associated with alternative optimization criteria and that these
externalities could have substantial impacts on our communities and on the nature of the routes we use. For instance, our
evidence suggests that scenic routing removes vehicles from highways (where city planners generally hope to route traffic)
and redirects them to parks, popular areas, and, in some cases, wealthier areas. Scenic routing also creates substantially more
complex routes involving more turns and intersections, both of which are known to make routes less desirable [14,24] and
are associated with negative outcomes (e.g. traffic accidents [33], decreased usability [36,56]). Additionally, safety routing
creates highly local but large impacts, redistributing traffic from pre-defined banned areas to highways and surrounding
thoroughfares.
Importantly, these externalities can arise even when increases in travel time appear minimal, providing evidence that
externalities may be transparent under the standard paradigm for evaluating alternative routing approaches. Along the same
lines, we also identified evidence that there is substantial variation in the externalities of a given algorithm across different
cities and areas within a city. As such, our findings suggest that alternative routing research must involve carefully
controlled evaluations across broadly diverse geographies to fully understand the costs and benefits of an algorithmic
change.
The algorithmic auditing component of our work reveals that Google Maps and MapQuest likely incorporate some non-
fastest-path optimizations in their routing algorithm (e.g. they generate simpler routes that spend more time on highways).
However, we found no evidence (yet) of any major externalities relative to fastest-path routing, indicating, for instance, that
Google Maps and MapQuest have not implemented features like Waze has in Rio de Janeiro and applied them at scale. Our

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:3
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
methods will allow us to easily monitor this result over time to assess if this changes (e.g. if one of these routing providers
begins to route people around neighborhoods with higher crime rates).
Finally, in the spirit of work such as Jurgens et al. [20], which also sought to standardize the evaluation of an algorithm
family (geolocation inference algorithms), we are releasing our alternative routing and evaluation platform to facilitate
improved evaluations and comparisons in this domain. In addition to allowing researchers to easily consider externalities
when evaluating new routing algorithms, our platform also addresses issues in the alternative routing literature such as a
lack of standards, very limited open alternative routing implementations, and inconsistent evaluation criteria. We have
designed our platform to be completely open-source and easily extensible (e.g., to other optimization criteria, geographic
scales, or externalities), with the hope of supporting future research and discussion of what is important in evaluating
geographic vehicle routing.
2 RELATED WORK
In this section, we discuss research that motivated this work. This research emerges primarily from four areas: the large
literature on alternative routing criteria, investigations into route preference, evaluation approaches in alternative routing,
and the algorithms underlying commercial mapping platforms. Notably, though the research in this paper focuses on vehicle
routing, we include in our discussion of related work approaches that have considered other modes of transportation as
well.
2.1 Routing Using Alternative Criteria
While there is a large and growing literature on developing alternatives to shortest and fastest path routing, there has not
yet been an effort to summarize this literature. As such, we conducted a survey of the literature and found that the alternative
routing approaches largely fall into four categories: positive, negative, topological, and personalized (see Table 1 below for
examples of each).
The first two categories, positive and negative, are defined by the work of Golledge [14], which examines in part the
impact of environmental features on route preferences (e.g. parks as positive, waste dumps as negative). The third category,
topological, encompasses criteria such as simplicity or driving efficiency that can be derived from basic information about
the road network. The final category is personalized routing, which involves learning the personal preferences of a driver
(e.g. road or turn types) and designing routes that adhere to these preferences.
Within the routing algorithms literature, positive routing often takes the form of “scenic” routing. Scenic routing has
been implemented in a number of ways, e.g. reweighting edges based on an assessment of the “scenicness” of their
surrounding area [55], adding waypoints from scenic areas near the shortest-path route [8,44,58], by generating many paths
and then choosing the most scenic [41]. Additionally, there are also examples of optimizing for other positive criteria such
as “happiness” and “quiet” [41] and projects peripheral to alternative routing that propose means of sensing criteria such as
desirable smells [42] for future use in routing.
Negative routing seeks to provide routes that avoid undesirable areas. Though Golledge used waste dumps as a proxy
for this type of routing, the literature largely focuses on avoiding unsafe areas as defined by high incidences of violent crime
[9,11,21,45]. Additional applications include avoiding dangerous weather [25] or other people [40]. Microsoft owns a patent
[51] for pedestrian routing that avoids unsafe areas. Waze has already included the option to avoid high-crime areas,
specifically in Rio de Janeiro, Brazil, ahead of the 2016 Olympics [39]. Waze also defaults to routing individuals around
Table 1. A selection of categorized alternative routing papers.
CATEGORY
SUB-CATEGORIES AND PAPERS
Positive
Scenic (El Ali et al. 2013, McGookin et al. 2015; Quercia, Schifanella, and Aiello 2014; Runge et al.
2016; Traunmueller et al. 2013; Zhang, Kawasaki, and Kawai 2008; Zheng et al. 2013)
Negative
High Crime (Elsmore et al. 2014; Fu, Lu, and Lu 2014; Kim, Cha, and Sandholm 2014; Shah et al.
2011), Weather (Y. Li et al. 2014), People (Posti et al. 2014)
Topological
Simplicity (Duckham and Kulik 2003; Haque, Kulik, and Klippel 2006; Shao et al. 2014), Health
(Sharker, Karimi, and Zgibor 2012), Efficiency (Ganti et al. 2010)
Personalized
Letchner, Krumm, and Horvitz 2006; Delling et al. 2015; Pang et al. 1995; Ziebart et al. 2008

15:4 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
certain settlements that are off-limits to Israelis [50]. These algorithms start with the shortest path and then either add
waypoints as needed to reroute away from any areas deemed undesirable or reweight edges in these areas so that they are
perceived as very high cost by the algorithm.
Topological routing describes approaches that seek to optimize on some aspect of the road network itself (rather than
the environment around the road network). The most common approach in the literature – other than the more traditional
travel time and distance criteria – is some form of “simplicity” routing. At their core, simplicity routing approaches seek to
model the ease with which a driver can follow a route, but they operationalize simplicity a number of different ways.
Golledge [14], Manley et al. [27], and Li and Wu [24] modeled simplicity as minimizing the total number of turns.
Algorithmic implementations of simplicity routing have taken the approach of modeling simplicity not just using turns, but
as a function of the degree of each intersection and what action is taken at that intersection (i.e. go straight or turn) [7,16,47].
Finally, personalized routing approaches generally seek to learn and model an individual’s route preferences, i.e. when a
user usually deviates from the fastest route and, in some cases, why s/he does so [5]. These models require extensive
positioning (i.e. “GPS”) data from drivers in order to determine these preferences. A common approach is to learn implicit
preferences for specific roads [23,60], though a recent approach by Delling et al. [5] explicitly learns the weights that each
individual driver appears to give to various topological criteria (e.g., number of lanes, type of road).
We included in our experiments the most common form of alternative routing approaches in each category, with the
exception of personalized routing. More specifically: for positive routing, we implemented scenic routing; for negative
routing, we implemented safety routing; and for topological routing, we implemented simplicity routing. We did not consider
personalized routing because the open “GPS” trace datasets that would be required to include personalized routing in our
experiments do not exist.
2.2 Preferences for Alternative Criteria
Researchers have long known that fastest path and shortest path are not the only criteria on which people want to optimize
their routes. Much of this knowledge emerges from a variety of surveys and field studies. For instance, in an influential
paper in the field of geography, Golledge [14] sought to quantify the importance of various criteria in route selection.
Golledge identified that minimizing distance and travel time were the most important factors, but minimizing the number
of turns and maximizing the scenic/aesthetic value were also key criteria that seemed to affect which route a participant
selected. Li and Wu [24] surveyed commuters in Florida and provided support for the findings of Golledge, but also
determined that safety is an important criterion. Similarly, Manley et al. [27] explored which criteria best explain actual
routes taken by taxi drivers in London and found that a combination of shortest distance and fewest turns was most
predictive of route choice. In addition to the preference for fewer turns, increased route complexity also raises safety
concerns [33] and has been directly tied to greater cognitive loads [36] and stress [56] for the driver. This literature, though
sparse, further motivates our choice of the three alternative criteria that we consider in this work (beauty, safety, and
simplicity).
2.3 Evaluation in Alternative Routing
As is often the case in new computing research areas (e.g. geolocation inference [20]), evaluation in the alternative routing
literature is a highly heterogeneous process that makes comparisons between approaches difficult. Evaluations typically
involve examining routing in 1-2 cities using a small number of routes, with the only evaluation metric being travel time or
distance. Our work is the first to our knowledge that explicitly considers additional evaluation criteria. In other words, this
work sheds new light on the externalities, or side-effects, that arise with the use of alternative routing optimization criteria.
Hints to the existence and importance of these externalities come from popular media, as discussed below. Like was the case
in Jurgens et al. [20] for geolocation inference, our goal in this paper is to conduct experiments that afford a more direct and
nuanced comparison between approaches, enabling a more robust understanding of the externalities associated with each
approach. We also examine routing in four cities with diverse geographic contexts, affording a broader view of how
geography and algorithms interact that provides important new insight.
2.4 Commercial Mapping Platforms
Since MapQuest began providing online directions in 1996, most online mapping platforms have defaulted to providing the
“fastest” route between a given origin and destination. The exact details of the routing algorithms being used are proprietary,

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:5
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
often including whether or not these algorithms are optimizing on criteria other than just travel time. Some information,
however, has been made public. Delling et al. [5] note that Microsoft Bing’s algorithm takes into account dozens of
topological features such as the type of road, number of lanes, speed limit, and historical traffic data, and that the algorithm
optimizes for simplicity as well by incorporating turn costs. Waze became infamous for especially accident-prone turns
across traffic (an externality that likely arose as a result of optimizing more heavily than other commercial routing platforms
on minimizing travel time) and has also since begun to explicitly optimize for simpler turns [39].
The introduction of the avoidance of “dangerous” areas in Rio de Janeiro by Waze [28] represents a large deviation from
the fastest route paradigm. Waze also has incorporated avoidance of areas that cannot be entered legally by certain
individuals, e.g. various settlements when driving in Israel [50]. Questions have been raised as to whether this type of safety
routing merely enforces stereotypes and unfairly removes traffic (including potential retail customers) from poorer areas
[28,39]. These concerns were also raised when a Microsoft patent that describes a means for helping pedestrians avoid areas
where crime has been reported became public [15,33].
The Waze platform has also been accused of routing its users through many previously low-traffic neighborhoods. This
has resulted in a number of externalities – including extensive frustration among residents (e.g. [18,29,57]) – and has led to
calls for regulation of where mapping platforms can direct traffic [29]. Relatedly, a simulation study found that near-
universal usage of fastest path routing during high-traffic times led to the redistribution traffic away from highways and
onto more local roads [52]. Our research extends our understanding of traffic redistribution externalities to the large
literature on alternative routing criteria and can help inform the policy debate about mapping platform regulation.
This research also builds on work that aims to provide some transparency to large-scale geographic systems that inform
how we interact with the world, such as that by Soeller et al. [49], who developed a system for detecting personalization of
political borders on Google Maps, and Chen et al. [4], who examined Uber surge pricing in San Francisco and Manhattan.
This research takes a similar approach, but moves towards detecting the employment of alternative routing criteria (e.g.
crime) rather than political borders or geographic biases in the sharing economy.
3 METHODS AND FRAMEWORK
In order to conduct a controlled evaluation of the externalities associated with alternative routing criteria, we needed three
main components: the routing algorithms for each alternative criterion (i.e. beauty, safety, simplicity), a set of origin and
destination pairs, and metrics to analyze the different externalities. For each origin-destination pair, and in aggregate across
all origin-destination pairs for a city, we are then able to directly compare the routes generated by each algorithm. Below,
we describe in detail our implementations of each of these components.
3.1 Alternative Routing Algorithms
We implemented three alternative routing approaches as well as a more traditional fastest-path algorithm to provide context
when necessary. For our alternative approaches, we selected scenic, safety, and simplicity routing. As noted in Section 2.1,
these three approaches have been validated by Golledge’s work and have been the subject of substantial interest in the
alternative routing algorithm literature. There is no consensus in this literature, however, on how to operationalize criteria
like “scenicness” (which is referred to as “beauty” in some literature), safety, or simplicity in a routing algorithm.
Additionally, there is also a lack of open implementations of these and other alternative routing approaches.
To address these issues, we developed our own framework that consists entirely of open-source software components
and publicly-available data. We have released this framework for others to use and improve2. As noted above, the primary
goals of the framework are (1) to make it easy for researchers and developers to consider externalities in their routing
algorithm evaluations and (2) to provide a greater degree of standardization in routing algorithm evaluation more generally.
Our framework is straightforward and has the benefit of being easily extensible to include additional alternative routing
approaches and additional externality metrics not discussed in this paper (e.g. number of stop lights [14] or the diversity of
neighborhoods along the route).
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
2 https://github.com/joh12041/route-externalities

15:6 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
For implementation of our alternative routing and fastest-path algorithms, our framework utilizes the bidirectional
Dijkstra implementation provided by the open-source GraphHopper Java library 3. GraphHopper includes standard
pathfinding algorithms and imports OpenStreetMap4 road networks to build the underlying graph for routing. GraphHopper
does not take traffic into account when determining the fastest path and instead bases travel time on road speed limits
included in the OpenStreetMap data. All of our alternative routing approaches are included in our released framework.
3.1.1 Scenic Routing. Our implementation of scenic routing is designed to replicate the approach described by Quercia
et al. [41], except where required by the nature of our study. Broadly, Quercia et al. collect the top k shortest paths between
two points and then select the path that optimizes for beauty based on data derived from Flickr photo tags. We chose this
approach because it is scalable to many geographic regions and complements open-source approaches as it relies on public
social media data.
Quercia et al. base their underlying data on a LIWC-based [37] text analysis of the tags on geotagged Flickr photos.
Specifically, they generate a 200m-by-200m grid across the city of interest in which each grid cell has a score based on its
corresponding geotagged photos tags. They validated this approach with crowdsourced ground-truth beauty rankings. The
notable variations from Quercia et al. in our system are as follows: our grid cells are slightly smaller and are not perfectly
square (0.001 degree by 0.001 degree to necessarily speed up the algorithm) and we use Empath (Fast et al. [10], a validated
open-source replacement for LIWC). In order to achieve better spatial coverage, we also add geotagged tweets5 to the Flickr
[54] tags that are used to score each grid cell. Despite these changes, as validation we note that we see similar trade-offs in
travel time to those reported by Quercia et al.
In addition to defining scenicness, the Quercia et al. approach also needs a means of generating and ranking routes based
on this alternative criterion. Quercia et al. use Eppstein’s algorithm, which finds the k-shortest paths between an origin and
destination. We do the same through GraphHopper, but find the k fastest paths where we cap k at 1,000 as Quercia et al.
demonstrated diminishing returns with larger k values (we also examined k=10,000 and found similar results to those
presented below, but with greater effect sizes). As is done in Quercia et al, each route is scored as the average beauty score
of the grid cells through which it passes. The route with the highest average beauty score is selected and returned.
3.1.2 Safety Routing. We implement safety routing as closely as possible to descriptions of the technique used by Waze
in Brazil. Though the specific details of the algorithm are not public, Waze notes that it avoids areas that have “higher-than-
average homicide, car robbery, or drug trafficking rates” [28]. We focus our efforts with respect to safety routing on New
York City6 and San Francisco7, both of which provide public crime data. We include data from all of 2016, retaining only the
crimes that overlap with the categories mentioned above. It was also noted that Waze disregarded areas that had high
numbers of drivers under the assumption that their users considered these areas to be safe [28]. Lacking this (private) data,
we implemented a proxy: we disregarded highways (speed limit greater than 70 kilometers per hour) when avoiding roads
in these areas.
Waze has not released the delineations of the areas in Rio de Janeiro that were designated “unsafe,” so we tested several
thresholds for determining which areas to instruct our algorithm to avoid. Waze chose 25 areas in Rio de Janeiro that are
described as varying in size between a block and a neighborhood [28]. As such, we aggregate the crime data to census tracts,
which in cities are generally of a size between a few blocks and a neighborhood. We then define a threshold for the average
number of crimes (normalized by the area of the census tract) such that a certain percentage of census tracts (i.e. those above
the threshold) are marked as “unsafe.” We tested different variations of our threshold such that it removes 25%, 15%, 10%,
5%, or 1% of census tracts8. We report results for the 10% threshold, finding the results for the other thresholds to be very
similar.
With a list of census tracts that exceeded the threshold for crime, we conducted safety routing in GraphHopper by using
fastest path routing but avoiding all road segments that pass through these census tracts and do not have a speed limit
greater than 70 kilometers per hour. Thus, a fastest path that does not pass through a blocked area will be unaffected while
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
3 https://github.com/graphhopper/graphhopper
4 https://www.openstreetmap.org
5 https://dev.twitter.com/streaming/overview
6 http://www.nyc.gov/html/nypd/html/analysis_and_planning/historical_nyc_crime_data.shtml
7 https://data.sfgov.org/
8 Actually operationalizing on all areas with “higher-than-average” crime would have blocked far too many census tracts – i.e. about a third of census tracts
in each city.

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:7
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
a path that would have passed through a blocked area is rerouted to the fastest path that does not pass through a blocked
area.
3.1.3 Simplicity Routing. We implemented simplicity routing as described by Shao et al. [47]. This approach scores the
simplicity of a route as the sum of the complexity of each intersection through which it passes. The complexity of an
intersection is modeled based on the degree of the intersection (i.e. number of intersecting roads) and what action is to be
taken by the driver (i.e. going straight or turning). We re-use the k-shortest-paths framework from scenic routing, but
instead of selecting the path with the highest average beauty, our simplicity algorithm selects the path that has the lowest
complexity score (i.e. the simplest route).
3.2 External APIs
We also included routes from two external mapping platforms9, Google and MapQuest, to address RQ3 (evidence of
alternative criteria in routes from third-party platforms). We report results for routes that were gathered on weekdays from
5-7:30pm local time for both platforms, a time of high traffic. We also gathered routes from 2-4:30am local time (weekdays)
for low traffic directions, but do not report these results as we found little difference in the actual routes (i.e. ~98% overlap
in routes, just different expected travel times).
3.3 Origin-Destination Pairs
In order to robustly compare the routes provided by different routing optimizations, we needed a set of origin-destination
coordinate pairs in each of our four cities. Ideally, analysis of routes would be done with a representative sample of route
requests (e.g. from Google Maps or MapQuest server logs). However, this type of data is not publicly available. To address
this issue, we take two approaches: (1) we adopt a common practice in the literature (e.g. [16,46,61]) and test the algorithms
on randomly-generated origin-destination pairs from across a city’s entire road network and (2) we also use publicly-
available taxi trip datasets where available.
With respect to our randomly-generated dataset, we generate approximately 5000 origin-destination pairs each for San
Francisco (California, USA), New York City (New York, USA), London (England), and Manila (Philippines). These cities were
chosen to provide regional variation while still having sufficient English speakers to provide a good source of photo tags
and tweets for our scenic routing algorithm (Empath is currently limited to English).
To provide additional context when possible, we verified the validity of these randomly-selected pairs by analyzing two
datasets of actual route origins and destinations based on taxi pick-ups and drop-offs, one in San Francisco [38] and the
other in New York City10. As we discuss below, for our route-level externalities (RQ1), we see the same high-level findings
in our taxi-sampled and randomly-generated datasets, and so we only report the results for the randomly-generated pairs.
For our community-level externalities (RQ2), we reach varying conclusions depending on whether we use the taxi-sampled
or randomly-generated origin-destination pairs. As such, we focus our discussion of RQ2 on San Francisco and New York
City and discuss both sets of results.
3.4 Externality Metrics
The first set of externalities that we examine are attributes of a route that are not traditionally considered in evaluations but
that have been found to be important in how people choose routes (i.e. RQ1, route-level externalities). First, we evaluate the
complexity of the route, which we measure in several ways: number of turns [14,26], number of left (or right in London)
turns [24], and the metric that we use in simplicity routing [7], which takes into account the number of intersections passed
through by a route and what action is taken at each intersection (i.e. turn or go straight). We also measure the beauty of all
of the routes [1,14,24] – another desired property of routes as determined by Golledge – doing so in the same way as
described above for our algorithmic implementation of scenic routing. Finally, due to the public outcry around Waze
redirecting traffic from the highways into smaller neighborhoods, we also measured the percentage of time that each route
was on highways (i.e. “motorways” in GraphHopper, which are operationalized as roads with speed limits greater than 70
kilometers per hour) and the percentage of time that each route was on slower neighborhood roads (i.e. streets below
“secondary” in GraphHopper, as operationalized as roads with speed limits less than or equal to 40 kilometers per hour). For
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
9 Waze does not provide a public API.
10 www.nyc.gov/html/tlc/html/about/trip_record_data.shtml

15:8 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
all of these metrics, we compute 99% confidence intervals by bootstrap resampling the routes 1000 times (i.e. sampling routes
with replacement from the approximately 5000 generated for each algorithm and city).
The second set of externalities relates to the community-level impact across all the routes considered (RQ2), with much
of the motivation for these externalities arising from the public discourse around Waze [28,29], i.e. analysis of how traffic
might be redistributed throughout a city and whether income appears to be a factor in this redistribution. We specifically
focus on income because concerns have been raised that safety routing would also lead to the avoidance of poorer
neighborhoods [15,28]. For these externalities, we focus on the cities in which we implemented safety routing (New York
City and San Francisco, both of which also have detailed census data on income available11). We compute how income
correlates with where an alternative routing algorithm redistributes traffic (as compared to traditional fastest-path
algorithms). Specifically, for road segments that saw significantly increased traffic for a given alternative routing algorithm,
we calculate the weighted average of household median income based on how much additional distance of roads passed
through a given census tract as compared to the GraphHopper fastest algorithm. For example, if across all origin-destination
pairs, there was an additional 8 km of routes in a census tract with a household median income of $60,000 and 2 km of routes
in a census tract with a household median income of $50,000, then the weighted average would be $58,000 for roads that
saw increased traffic. We do the same then for road segments that saw significantly less traffic and compare. We again
compute 99% confidence intervals through bootstrap resampling with 1000 iterations on the origin-destination pairs.
3.5 Calculating Metrics for Commercial Mapping Platforms
The MapQuest and Google APIs provide the points that comprise the route that they return (i.e. latitude, longitude of enough
points to accurately convey the geometry of the route). From these points, we can easily calculate the beauty of the MapQuest
and Google routes through the same beauty grid-cell framework as used with GraphHopper. However, calculating the
simplicity for these routes is less straightforward because the details of each intersection are not provided by Google and
MapQuest. To overcome this problem, we use the map matching process developed by Newson and Krumm [34], which has
also been implemented in GraphHopper12. This process converts the points into a corresponding path on the GraphHopper
road network, from which simplicity can then be calculated as before. The Newsom and Krumm approach is not perfectly
accurate, however, and so we enforce that the resulting matched route must be within 5% of the length of the original route
in order to be considered in further analyses. Recall is generally around 85%, with the exception of Manila at 63%, which
likely reflects differences in the underlying road networks of OpenStreetMap and the commercial mapping platforms.
4 RESULTS
In this section, we analyze and compare the routes (i.e. Google Fastest, MapQuest Fastest, GraphHopper Fastest,
GraphHopper Scenic, GraphHopper Safe, GraphHopper Simple) according to the metrics described in the prior section.
4.1 RQ1: Route-Level Externalities for Alternative Routing Approaches
The routes corresponding to the San Francisco origin-destination pair featured in Figure 1 are illustrative of the type of
route-level externalities that are seen between the different optimizations. In general, we see that the Google, MapQuest,
and GraphHopper Simple paths share many of the same characteristics and are longer than the GraphHopper Fastest path,
making more extensive use of highways. The GraphHopper Scenic route looks quite different from both the simplest and
fastest routes, taking a more complicated path that passes through several popular areas such as Union Square before
arriving at the destination. The GraphHopper Safe path also deviates substantially from the fastest path in order to
circumvent the Tenderloin, an area of higher crime in San Francisco.
These trends in the route-level externalities that arise as a result of different optimization criteria can be seen in Figure
2, which shows the results of each route-level evaluation metric by Euclidean distance between the origin and destination
(x-axes) and city (columns). We walk through Figure 2 in the sub-sections below and supplement the trends with statistics
from Table 5 in the Appendix, which contains the actual values for each city and algorithm when the Euclidean distance is
between 10 and 11 kilometers (i.e. one slice of the data in Figure 2). Of note, we present the externalities both normalized to
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
11 https://factfinder.census.gov/faces/nav/jsf/pages/index.xhtml
12 https://github.com/graphhopper/map-matching

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:9
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
the natural baseline (e.g. GraphHopper Simple for simplicity measures) and also in absolute units when the units are readily
interpretable (e.g. # of turns).
4.1.1 Route Complexity. Optimizing on beauty or safety substantially increases the complexity of routes, which has
implications for driver safety and usability. As can be seen in rows 1 and 2 of Figure 2, the average number of turns and
therefore complexity of a route increases significantly when comparing the fastest path to either the scenic or safer paths.
Specifically, for route distances of 10-11 kilometers, the scenic path takes an additional 4-5 turns over the fastest path and
the safer path takes on average an additional turn. Similar trends hold for the number of left turns (right turns in London)
on a route (not shown) and the simplicity of a route (not shown, as measured by the number of intersections and action
taken at each intersection). The simplest path is then an additional 2-3 turns shorter than the fastest path. For the safest
path, this increase in complexity is the result of the added distance to circumvent a given area, but, for scenic routes, there
are also significantly more steps per kilometer as well (third row of Figure 2).
4.1.2 Beauty. The scenic route is about 2-3x more beautiful (row 4 of Figure 2) than the other routes produced with other
optimization criteria depending on the city and Euclidean distance. Notably, neither simplicity nor safety seems to correlate
significantly with increased or decreased beauty.
4.1.3 Time on Highways and in Neighborhoods. Given concerns about the shift of vehicles away from highways and onto
smaller neighborhood roads, rows 5 and 6 in Figure 2 consider the time spent by each route on each type of road. Scenic
routes spend proportionally less time on the highway than the Google, MapQuest, or GraphHopper Fastest routes while
GraphHopper Simple routes spend proportionally more time on the highway. Intuitively, this makes sense – many of the
scenic spots in cities are not next to highways and taking a highway generally limits the number of intersections
encountered. The magnitude of the differences varies across cities. On the low end, the scenic routes in London spend about
0.5% less of their time on the highway than the fastest path. At the high end, in San Francisco, the scenic routes spend about
9% less of the route on the highway than the fastest path, which corresponds to a 70% relative decrease in the amount of
time spent on highways.
Scenic routes spend a significantly greater proportion of their travel time on slower roads, i.e. smaller roads that generally
go through residential neighborhoods, foreshadowing their community-level effects highlighted below. The specific
proportion of time spent on these roads varies greatly by city, but the GraphHopper Fastest, Simple, and Safe routes on
average spend similar proportions of time on these roads whereas scenic routes tend to spend 25-50% (relative) more of their
travel time on these roads.
!
Figure 1. Routes given by each routing algorithm for an example origin-destination pair in San Francisco.

15:10 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
!
!
Figure 2. Comparison of the route-level externalities (rows) by city (columns), distance between the origin and
destination pair (x-axis), and routing algorithm (lines). 99% confidence intervals calculated through bootstrap
resampling. Average travel time (bottom row) for Google Fastest and MapQuest Fastest routes are not shown because
any differences between them and the GraphHopper routes likely arise from variation in the underlying data for
calculating travel time.

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:11
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
4.2 RQ2: Community-Level Externalities of Alternative Routing Approaches
Motivated by public concern around the redistribution of traffic and disproportionate impacts on poor (or wealthy) areas,
we also examined community-level externalities that may arise from optimizing on alternative criteria. Examples of these
externalities are visualized in Figures 3 and 4, which show the roads that see significantly more or less traffic in New York
City and San Francisco for scenic, safety, and simplicity routing. We show both the results from the randomly-generated
origin-destination pairs, which provide good coverage of the entire areas, and the taxi-sampled origin-destination pairs,
which are more representative of actual route concentrations.
4.2.1 Distribution of Traffic. For scenic routing, areas around parks see greatly increased traffic, as do popular tourist
destinations and commercial districts. As can be seen in Figure 4, in San Francisco, the largest increases (>100 additional
Figure 3. Comparison of alternative routing algorithms and GraphHopper Fastest algorithm, showing road segments
with a significant change in the number of routes that pass over them across all origin-destination pairs for
GraphHopper Scenic (left), Safe (middle), and Simple (right) routing in New York City. Results for random origin-
destination pairs are shown on top and taxi-sampled origin-destinations are shown on bottom. Blue road segments
had more routes pass over them with the alternative routing algo rithm as compared to the Gra phHopper Fastest
algorithm, and red road segments had fewer routes pass over them with the alternative routing algorithm as
compared to GraphHopper Fastest. The specific color thresholds were set by quantiles with the constraint mentioned
above that blue represents increased traffic and red represents decreased traffic. Darker colors indicate a greater
magnitude in the difference in number of routes passing over a given road segment. Black slanted lines in the
GraphHopper Safe maps indicate blocked areas in safety routing.

15:12 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
routes out of the approximately 5000 analyzed in the random origin-destination pairs) occur around Golden Gate Park, the
Embarcadero (popular waterfront region), Glen Park, and Mission Street as it passes through the Mission District (popular
commercial district). The roads that see corresponding drops in traffic are often nearby highways or similarly large
thoroughfares, which have high speed limits but are not always scenic. In New York City (Figure 3), scenic routing led to
the largest increases in traffic (again >100 out of the approximately 5000 routes in the random origin-destination pairs) on
roads that border the rivers, the roads around Central Park, and in Williamsburg (a rapidly gentrifying neighborhood in
Brooklyn). Again, it is largely highways and nearby thoroughfares where the greatest decreases in traffic are seen.
The changes in traffic related to safety routing are much more localized, with increased traffic in order to circumvent the
blocked census tracts being redirected to highways as well as more local roads that are immediately outside of the blocked
areas. In San Francisco, the region that sees the largest decrease in traffic is the Tenderloin (a poorer neighborhood very
close to downtown). In New York City, the taxi-generated routes show that the bulk of the traffic redistribution would occur
to avoid high-crime areas in Manhattan, though the randomly-generated pairs indicate that regions of Brooklyn and Harlem
would see traffic shifted to the highways as well.
Simplicity routing leads to a large increase in the amount of traffic on highways, as they have fewer intersections, but
does not appear to favor or avoid any specific areas.
4.2.2 Income of Neighborhoods. Given concerns about safety routing criteria avoiding poorer neighborhoods, we also
computed the weighted average of the household median income for roads that saw significantly increased or decreased
traffic. The results are provided in Table 2.
Figure 4. Comparison of alternative routing algorithms and GraphHopper Fastest algorithm, showing road segments
with a significant change in the number of routes that pass over them across all origin-destination pairs for
GraphHopper Scenic (left), Safe (middle), and Simple (right) routing in San Francisco. Results for random origin-
destination pairs are shown on top and taxi-sampled origin-destinations are shown on bottom. Blue road segments
had more routes pass over them with the alternative routing algo rithm as compared to the GraphHopper Fastest
algorithm, and red road segments had fewer routes pass over them with the alternative routing algorithm as
compared to GraphHopper Fastest. The specific color thresholds were set by quantiles with the constraint mentioned
above that blue represents increased traffic and red represents decreased traffic. Darker colors indicate a greater
magnitude in the difference in number of routes passing over a given road segment. Black slanted lines in the
GraphHopper Safe maps indicate blocked areas in safety routing.

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:13
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
Across the different algorithms and cities, we see mixed but persuasive results that alternative optimization can lead to
large and disparate externalities in the types of areas that receive increased or decreased traffic. Scenic routing favors
wealthier areas in both cities. The taxi-sampled (and arguably therefore more representative of actual traffic patterns) origin-
destination pairs indicate that traffic on average moves towards wealthier regions - i.e. the average household median
income of areas that see increased traffic is significantly higher than that of areas that see decreased traffic. For instance, in
San Francisco for the taxi-sampled origin-destination pairs, the household median income of road segments that saw
increased traffic was $97,639 while it was only $92,135 for road segments that saw decreased traffic. The randomly-sampled
pairs in New York City indicate no significant correlation with traffic changes and income, though scenic routing causes
traffic to move to less wealthy areas in San Francisco when using the randomly-sampled pairs.
Safety routing results in mixed effects across the two cities. In San Francisco, for both the taxi-sampled and randomly-
generated origin-destination pairs, we see that safety routing moves traffic towards wealthier areas. The average household
median income of areas that see increased traffic is $15,000 higher than that of the areas that see decreased traffic. In New
York City, we see smaller disparities in the income of areas where traffic is redistributed, but the safety routing approach
actually seems to cause traffic to shift on average towards less wealthy areas. Adding to the robustness of these results, we
note that we see the same trends if we look at alternative metrics such as the proportion of increased and decreased traffic
in areas with a household median income below a given threshold, e.g. $40,000.
4.3 RQ3: Alternative Criteria in External Mapping Platforms
Our results suggest that Google and MapQuest are likely incorporating simplicity as an optimization criterion in addition to
travel time (like Bing [5]), generating simpler routes than would be expected under a pure fastest-path approach. For
instance, Figure 2 shows that both Google and MapQuest provide routes that are similar to the GraphHopper Fastest and
GraphHopper Simple routes. Interestingly, we calculated the percentage overlap between each combination of the various
GraphHopper and external platform routes and found that MapQuest and Google are most similar to each other but that the
highest overlap between Google or MapQuest and the GraphHopper routes is with GraphHopper Simple and not with
GraphHopper Fastest.
Importantly, however, we do not see evidence of Waze-style safety routing being applied in either platform, i.e. neither
Google nor MapQuest appear to be excluding any neighborhoods from their routes. More generally, we also observe no
major externalities relative to GraphHopper Fastest in either commercial platform, aside from an increase in simplicity. This
can be seen in Figure 5, which shows the significant differences (at 99% confidence) in the number of routes that pass over
a given road segment when comparing Google and MapQuest Fastest versus GraphHopper Fastest. While many roads show
different levels of traffic, there is no clear geographic concentration in roads that are favored or avoided. Looking back at
Figures 3 and 4, we see that scenic and safety routing resulted in areas in the city where many roads all saw an increase (e.g.
a popular and pretty neighborhood in scenic routing) or a decrease in traffic (e.g. an “unsafe” area in safety routing). We do
not see these patterns appear for Google or MapQuest in Figure 5.
Table 2. Average household median income of road segments that see significantly increased or decreased traffic with
each alternative optimization criteria. Using origin-destination pairs derived from taxi routes (i.e. reflective of actual
travel patterns as opposed to randomly generated) often demonstrates a higher income disparity between the types of
roads preferred or avoided by a given alternative optimization. City-specific differences are evident as well.
Origin-Dest.
Pairs
Change in
Traffic
Household Median Income [99% Confidence Interval]
Scenic
Safe
Simple
New York City
(Random)
Increase
$56,885 [55,484-57,555]
$59,110 [59,076-59,379]
$61,881 [61,838-62,160]
Decrease
$55,902 [55,745-56,233]
$60,338 [59,256-62,561]
$59,283 [58,992-59,961]
New York City
(Taxi)
Increase
$91,737 [90,997-92,590]
$91,870 [91,485-92,505]
$77,527 [74,021-78,618]
Decrease
$88,209 [87,377-89,607]
$98,779 [96,982-101,834]
$89,106 [87,877-91,635]
San Francisco
(Random)
Increase
$93,579 [93,024-93,891]
$87,352 [86,844-87,498]
$94,203 [93,868-94,383]
Decrease
$99,039 [98,505-100,387]
$72,566 [70,590-73,357]
$98,315 [98,002-98,858]
San Francisco
(Taxi)
Increase
$97,639 [97,077-98,528]
$76,660 [74,690-77,841]
$87,688 [87,543-88,184]
Decrease
$92,135 [91,147-94,439]
$59,279 [56,607-60,863]
$73,772 [70,242-75,430]

15:14 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
Figure 5. Comparison of third-party mapping platforms and GraphHopper Fastest algorithm, showing road segments
with a significant change in the number of routes that pass over them across all origin-destination pairs for Google
(left) and MapQuest (right) routing in San Francisco (top) and New York City (bottom). Results are shown for random
origin-destination pairs. Blue road segments had more routes pass over them with the alternative routing algorithm
as compared to the GraphHopper Fastest algorithm, and red road segments had fewer routes pass over them with the
alternative routing algorithm as compared to GraphHopper Fastest. The specific color thresholds were set by
quantiles with the constraint mentioned above that blue represents increased traffic and red represents decreased
traffic. Darker colors indicate a greater magnitude in the difference in number of routes passing over a given road
segment.

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:15
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
5 DISCUSSION
In this section, we discuss the implications of the above findings for both the design of routing algorithms and for public
policy.
5.1 Societal Impacts of Alternative Routing
A clear high-level finding in the above results is that alternative optimization criteria are associated with important
externalities that have not been previously considered. For instance, our results suggest that scenic routing (and safety
routing to a lesser degree) led to substantially more complex routes involving several more turns on average. Turning, and
specifically turns against traffic, are known predictors of collisions [30] and are less preferred by users [14]. Additional turns
also present a usability challenge, with more complex routes leading to greater cognitive load [36], increased driver stress
[56], and a greater likelihood of wrong turns and increased driving distance [16].
We also found that, if widely deployed, alternative optimization criteria such as beauty and safety would very likely lead
to some of the externalities that have been a matter of public discourse and frustration with Waze [18,28,29,57]. Scenic
routing redistributes traffic from highways into parks, popular areas, and onto slower, neighborhood roads. This raises
concerns that optimizing on beauty could further contribute to the frustrations about increased traffic in previously low-
trafficked neighborhoods [29]. These traffic increases on local roads have already led to calls by city council representatives
in several cities to limit where mapping applications can direct their users [18]. Alongside the frustration and potential
safety concerns of residents, high levels of traffic have also been tied to negative health outcomes [35].
By design, current safety routing approaches remove traffic from specific communities, which clearly could lead to
economic and social impacts on those communities. While important to recognize, that this occurs in safety routing is not
surprising. More surprising, however, is that traffic is not always just redistributed to surrounding (and likely similar)
communities, but instead has far-reaching impacts and often is moved to highways and major thoroughfares that circumvent
the larger area. Just as concerns have been raised about filter bubbles associated with the personalization of information
consumption (e.g. [22,49]), safety routing may perform a similar function, allowing people to avoid areas that they do not
want to see [28] and potentially shaping how we perceive the world [12]. A more balanced approach to safety routing might
implement the avoidance of these areas only at times of low traffic and, during high-traffic times of the day, actually favor
the “dangerous” areas so as to reduce the potential economic and social impacts of decreased traffic and visibility of these
areas. Additionally, rather than focusing on high rates of crime, safety routing might instead focus on reducing the risk of
collision by avoiding more dangerous driving maneuvers or crowded areas where accidents are more likely to occur.
In this paper, we explored previously-proposed alternative routing criteria with the concern that these could lead to
adverse and disparate impacts in specific areas. One can also imagine, though, alternative routing criteria that lead to a
greater diversity of experiences for the driver and more uniform impact on neighborhoods. We invite further discussion of
what other alternative criteria might be considered that would arguably lead to positive externalities.
5.2 Towards Improved Routing Evaluations
Our results also suggest that traditional routing algorithm evaluations are insufficient to capture the potential for
externalities. Specifically, the sole focus of traditional routing algorithm evaluations has been on potential increases in travel-
time or distance, but this can hide important negative outcomes such as increased complexity (and its associated safety
effects) and undesirable traffic patterns. Additionally, focusing just on travel-time and distance can lead to the conclusion
that the trade-offs of an alternative optimization diminish rapidly with distance whereas we find externalities whose effects
are relatively constant across distance. As can be seen in the final row of Figure 2, the increased travel time costs for the
alternative optimizations diminish as the route distance increases (this matches what is seen by Quercia et al. [41] as well).
However, this drop-off in magnitude of the trade-off is not nearly as immediate or does not occur when examining certain
externalities, such as the number of turns or what types of roads comprise the routes.
5.3 Algorithmic Auditing
In this paper, we provided some of the first audits of routes generated by major mapping platforms. We did not find any
evidence of major negative externalities associated with Google and MapQuest routes, with values for the externalities that
we studied generally falling in the same range as those for GraphHopper Fastest and GraphHopper Simple.

15:16 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
As more commercial mapping platforms take steps like Waze has done to include notions of safety or other alternative
optimizations in their algorithm, it is important that the research community continue to analyze the routes that these
platforms are providing to ensure that any negative externalities are monitored. This is especially critical given the extent
to which these platforms increasingly define the movement patterns of millions of people around the world. We have built
our platform to accommodate the evaluation of routes from any external source by incorporating the map-matching
component that converts a series of latitude-longitude coordinates to a route on the OpenStreetMap-based road network
used internally by GraphHopper. Furthermore, the dataset of routes collected in the course of this research can serve as a
baseline for future evaluations in order to detect changes in the routes provided by commercial mapping platforms.
5.4 Geography and Algorithms
In this research, we found that the externalities associated with a given routing algorithm varied across our four cities. For
route-level externalities, we saw different effect sizes in each city, but the broad trends remained consistent. Among
community-level externalities though, the nature of the trends changed from city to city. For instance, safety routing
redistributed traffic to less wealthy areas in New York City and to more wealthy areas in San Francisco. The dependence of
externalities on local geography extended to our choice of origin-destination pairs as well – i.e. randomly-generated vs.
sampled from taxi routes. For route-level externalities, the choice of origin-destination pairs did not affect the trends, but
the different sets of origin-destination pairs did lead to different conclusions for our examination of community-level
externalities.
These results indicate that the interaction between routing algorithms and geography, especially when evaluating
community-level effects, is highly dependent on the underlying urban structure and on origin-destination patterns within
that structure. Care should be taken when generalizing results from one or two cities to other settings. Performance and
results that vary across geographic contexts (“geographic human-computer interaction” [17]) has been highlighted in other
domains as well – e.g., highly skewed performance of geolocation algorithms in urban versus rural areas [19], decreased
precision predicting human perceptions of urban landscapes in a city when images and labels from a more distant city were
used to train the model [32], different levels of anonymity in location-based social networks in different regions [43], and
varying effectiveness of the sharing economy depending on the socioeconomic status of a given area [53]. Especially given
the ubiquity of these algorithms (e.g. Google Maps alone has over a billion unique monthly users [6]), expanding alternative
routing research to incorporate more geographic contexts will be important for guiding the design of these algorithms and
supporting continued public discourse. We hope that our evaluation platform can assist in this endeavor.
6 FUTURE WORK AND LIMITATIONS
One of the large questions raised by this work is how might we design alternative routing algorithms in such a way as to
realize their promised benefits while reducing the associated negative externalities. The literature provides some hints that
are worth exploring. For instance, the simplicity routing literature notes that hybrid, multi-criteria approaches (e.g.
balancing how much weight is given to simplicity and how much is given to travel time) often greatly reduce the complexity
of routes while incurring minimal time costs [16,47]. Further study could examine whether multi-criteria optimizations, or
other approaches that might directly consider externalities as a cost, show promise for reducing externalities. While we
mentioned one way in which safety routing might be implemented in a more balanced form in Section 5.1, further insight
might be gleaned from the urban studies literature (e.g. Jane Jacobs).
While the public discourse around Waze inspired some of this work, Waze currently does not provide a public API.
Future work might pursue alternative means of collecting Waze routes, in part as an automatic means of detecting whether
Waze routes are avoiding new areas. Furthermore, future work could also take advantage of open-source traffic data (e.g.
[31]) to better understand the behavior of commercial mapping platform routes and explore how alternative routing
algorithms react to changes in traffic as well.
We chose four cities as our geographic context for this study and sought to include cities from around the world.
However, as noted in the discussion of geography and algorithms in Section 5.4, it is very probable that different impacts
would be seen in other geographic contexts, such as in new cities or in suburban and rural areas. An interesting line of work
would involve categorizing different cities based on how these algorithms perform so as to build a better understanding of
how to more effectively target areas for study. In other words, are there classes of urban structures in which routing

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:17
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
algorithms tend to have similar externalities? Of course, repeating our research in suburban and rural areas is also an
important direction of future work.
We view this research as a first step towards understanding the externalities associated with various routing criteria. We
sought to design our alternative routing algorithms based on actively-developed open-source libraries (e.g. GraphHopper,
map matching) or published and validated methods (e.g. Empath [10], the k-shortest path approach developed by Quercia et
al. [41]). However, there are many parameters and possible approaches to alternative routing, which, if executed differently,
might lead to different results. Similarly, we analyzed the routes provided by Google and MapQuest for an initial directions
request, but it is possible that these routes would be updated as they are driven in order to take advantage of shortcuts off
of the highways. Finally, we built on previously-published methods to generate our alternative routes, but incorporating in
human assessments of the resultant routes would provide additional certainty that the routes would be perceived as more
scenic or simpler or safer.
7 CONCLUSION
In this paper, we provide the first robust assessment of the externalities associated with different alternative optimization
criteria in geographic vehicle routing. We show that these externalities are substantial and would not be detected by
traditional routing evaluations. For instance, we find that scenic and safety routing lead to more complex routes – increasing
accident risks and other negative effects – as well as substantially increased traffic in various communities. The community-
level impacts vary across different cities, however, demonstrating that evaluation across multiple geographic contexts is
necessary in order to understand the impact of alternative routing approaches and highlighting the complex relationship
between geography and algorithms more generally. We do not find evidence of negatives externalities arising in Google and
MapQuest but release our evaluation platform so as to support continued evaluation and monitoring of commercial routing
platforms. Finally, we discuss how algorithm designers might better balance the benefits of alternative optimization criteria
with the externalities that can arise through their use.
ACKNOWLEDGMENTS
The authors would like to thank our colleagues in GroupLens Research at the University of Minnesota for their help in
conceiving this project, especially Loren Terveen. This research was supported in part by the National Science Foundation
(CAREER IIS-1707296 and IIS-1707319) and a Lichtenberg Professorship from the Volkswagen Foundation.
!

15:18 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
A ADDITIONAL TABLES AND FIGURES
Table 5. Rankings of the optimization criteria as well as values and associated 99% confidence intervals for origin-
destination pairs separated by a Euclidean distance of 10 kilometers (i.e. numbers shown by the lines in Figure 2 when
straight-line distance = 10). Thou gh the magnitudes vary city to city, the ra nkings are largely stable across cities.
Abbreviations: GHSc = GraphHopper Scenic; GHSi = GraphHopper Simple; GHSa = Gra phHopper Safe; GHF =
GraphHopper Fastest; GF = Google Fastest; MF = MapQuest Fastest.
Metric
Value [99% Confidence Interval] for origin-destination pairs with a Euclidean-Distance of 10km
London
Manila
San Francisco
New York City
# of Turns
GHSc
MF
GHF
GF
GHSi
20.0 [19.0-20.9]
14.4 [13.6-15.3]
14.2 [13.2-15.2]
14.2 [13.1-15.2]
11.6 [10.9-12.3]
GHSc
MF
GF
GHF
GHSi
18.8 [17.4-20.0]
13.6 [12.7-14.7]
11.5 [10.3-12.7]
11.0 [10.1-12.1]
9.36 [8.54-10.2]
GHSc
MF
GHSa
GHF
GF
GHSi
14.0 [13.6-14.5]
12.0 [11.6-12.5]
11.3 [10.9-11.8]
10.6 [10.2-11.0]
10.4 [10.0-10.8]
7.67 [7.36-7.96]
GHSc
MF
GF
GHSa
GHF
GHSi
12.7 [12.2-13.3]
12.4 [11.9-12.8]
11.4 [10.9-12.2]
10.1 [9.48-10.6]
8.80 [8.32-9.30]
6.80 [6.47-7.12]
# of Turns per
kilometer
GHSc
GHF
MF
GF
GHSi
1.37 [1.30-1.46]
1.12 [1.04-1.20]
1.09 [1.03-1.15]
1.06 [0.99-1.13]
0.79 [0.84-0.89]
GHSc
MF
GHF
GF
GHSi
1.18 [1.10-1.25]
0.98 [0.91-1.05]
0.81 [0.73-0.88]
0.77 [0.69-0.85]
0.57 [0.63-0.69]
GHSc
MF
GHSa
GHF
GF
GHSi
1.07 [1.04-1.11]
0.88 [0.85-0.92]
0.86 [0.83-0.90]
0.82 [0.79-0.86]
0.73 [0.70-0.76]
0.56 [0.54-0.59]
GHSc
MF
GF
GHSa
GHF
GHSi
0.92 [0.88-0.97]
0.91 [0.88-0.95]
0.85 [0.80-0.90]
0.79 [0.74-0.83]
0.70 [0.66-0.74]
0.49 [0.46-0.51]
Simplicity
(Normalized)
GHSc
GF
GHF
MF
GHSi
1.46 [1.41-1.52]
1.17 [1.11-1.24]
1.14 [1.09-1.20]
1.07 [1.02-1.12]
1.00 [0.95-1.04]
GHSc
GHF
GF
MF
GHSi
1.71 [1.60-1.81]
1.16 [1.09-1.25]
1.13 [1.05-1.21]
1.12 [1.04-1.20]
1.00 [0.94-1.07]
GHSc
GHSa
GHF
GF
MF
GHSi
1.49 [1.46-1.52]
1.28 [1.23-1.32]
1.24 [1.20-1.29]
1.08 [1.05-1.11]
1.04 [1.01-1.07]
1.00 [0.97-1.03]
GHSc
GHSa
GHF
GF
MF
GHSi
1.54 [1.49-1.58]
1.30 [1.25-1.36]
1.24 [1.19-1.29]
1.20 [1.14-1.26]
1.08 [1.04-1.13]
1.00 [0.96-1.04]
Beauty
(Normalized)
GHSc
MF
GHF
GHSi
GF
1.00 [0.94-1.06]
0.36 [0.30-0.43]
0.34 [0.30-0.39]
0.34 [0.29-0.39]
0.32 [0.26-0.37]
GHSc
GHF
GHSi
GF
MF
1.00 [0.95-1.05]
0.41 [0.37-0.46]
0.40 [0.35-0.45]
0.38 [0.33-0.44]
0.38 [0.33-0.43]
GHSc
GF
MF
GHSi
GHSa
GHF
1.00 [0.97-1.04]
0.53 [0.49-0.58]
0.53 [0.49-0.58]
0.50 [0.46-0.54]
0.46 [0.43-0.49]
0.45 [0.42-0.48]
GHSc
GHSi
GHF
GHSa
MF
GF
1.00 [0.95-1.05]
0.39 [0.35-0.44]
0.39 [0.35-0.43]
0.38 [0.34-0.42]
0.34 [0.30-0.38]
0.31 [0.27-0.36]
% Time Spent
on Roads > 45
mph
MF
GF
GHSi
GHF
GHSc
2.84 [1.24-4.74]
2.76 [1.18-4.91]
2.54 [1.18-4.32]
2.07 [0.74-3.73]
1.50 [0.63-2.51]
GHSi
GF
GHF
MF
GHSc
17.5 [13.1-22.5]
13.5 [9.56-17.4]
11.9 [8.31-16.3]
10.9 [7.70-14.6]
6.18 [4.08-8.54]
GHSi
MF
GHF
GF
GHSa
GHSc
15.0 [13.0-17.1]
14.0 [12.2-16.1]
13.5 [11.7-15.4]
13.5 [11.3-15.6]
13.3 [11.1-15.5]
4.24 [3.19-5.41]
GHSi
MF
GF
GHF
GHSa
GHSc
27.8 [24.7-31.0]
23.6 [20.5-26.4]
21.4 [18.8-24.5]
20.5 [17.6-23.7]
20.5 [17.5-23.6]
15.3 [13.0-18.0]
% Time Spent
on Roads <
25mph
GF
GHSc
MF
GHSi
GHF
26.7 [23.0-30.4]
21.8 [19.2-24.8]
17.3 [14.4-20.4]
16.4 [13.6-19.2]
14.8 [12.4-17.6]
GHSc
GF
GHSi
MF
GHF
17.3 [14.8-19.9]
17.1 [14.3-20.1]
14.1 [11.1-17.4]
13.1 [10.6-16.0]
11.5 [9.23-13.9]
GF
GHSc
MF
GHF
GHSa
GHSi
26.7 [24.5-29.0]
24.2 [22.6-25.7]
23.3 [21.2-25.3]
16.1 [15.0-17.4]
16.1 [14.9-17.4]
16.0 [14.6-17.5]
GF
MF
GHSc
GHSi
GHSa
GHF
30.6 [27.8-33.5]
18.4 [16.6-20.4]
15.7 [14.5-17.0]
12.9 [11.5-14.4]
12.1 [10.9-13.4]
9.95 [8.89-11.1]
!
Table 6. Total number of user-generated contributions (i.e. photos, tweets) for each city used in determination of
beauty scores for scenic routing. Flickr photos from YFCC100M and Twitter tweets collected from May-August 2015.
Dataset
London
Manila
San Francisco
New York City
Flickr [# photos]
1,320,553
29,530
776,790
1,127,440
Twitter [# tweets]
262,216
328,179
200,609
685,172

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:19
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
REFERENCES
[1]
Majid Alivand, Hartwig Hochmair, and Sivaramakrishnan Srinivasan. 2015. Analyzing how travelers choose scenic routes using route
choice models. Computers, Environment and Urban Systems 50: 41–52.
[2]
Mike Ananny, Karrie Karahalios, Christian Sandvig, and Christo Wilson. 2015. Auditing Algorithms from the Outside: Methods and
Implications. In ICWSM.
[3]
Hannah Bast, Daniel Delling, Andrew Goldberg, Matthias Müller-Hannemann, Thomas Pajor, Peter Sanders, Dorothea Wagner, and
Renato F. Werneck. 2015. Route planning in transportation networks. arXiv preprint arXiv:1504.05140.
[4]
Le Chen, Alan Mislove, and Christo Wilson. 2015. Peeking Beneath the Hood of Uber. In IMC. https://doi.org/10.1145/2815675.2815681
[5]
Daniel Delling, Andrew V. Goldberg, Moises Goldszmidt, John Krumm, Kunal Talwar, and Renato F. Werneck. 2015. Navigation made
personal: inferring driving preferences from GPS traces. 1–9. https://doi.org/10.1145/2820783.2820808
[6]
Jillian D’Onfro. 2015. Here are all the Google services with more than a billion users. Business Insider. Retrieved from
http://www.businessinsider.com/google-services-with-1-billion-users-2015-10
[7]
Matt Duckham and Lars Kulik. 2003. “Simplest” Paths: Automated Route Selection for Navigation. In International Conference on Spatial
Information Theory, 169–185.
[8]
Abdallah El Ali, Sicco NA Van Sas, and Frank Nack. 2013. Photographer paths: sequence alignment of geotagged photos for
exploration-based route planning. In Proceedings of the 2013 conference on Computer supported cooperative work, 985–994. Retrieved
April 4, 2017 from http://dl.acm.org/citation.cfm?id=2441888
[9]
Samuel Elsmore, Irwan Fario Subastian, Flora Dilys Salim, and Margaret Hamilton. 2014. VDIM: Vector-based Diffusion and
Interpolation Matrix for Computing Region-based Crowdsourced Ratings: Towards Safe Route Selection for Human Navigation. In
MUM (MUM ’14), 212–215.
[10]
Ethan Fast, Binbin Chen, and Michael S. Bernstein. 2016. Empath: Understanding Topic Signals in Large-Scale Text. 4647–4657.
https://doi.org/10.1145/2858036.2858535
[11]
Kaiqun Fu, Yen-Cheng Lu, and Chang-Tien Lu. 2014. TREADS: A Safe Route Recommender Using Social Media Mining and Text
Summarization. In SIGSPATIAL (SIGSPATIAL ’14), 557–560.
[12]
Birgitta Gatersleben, Niamh Murtagh, and Emma White. 2013. Hoody, goody or buddy? How travel mode affects social perceptions in
urban neighbourhoods. Transportation research part F: traffic psychology and behaviour 21: 219–230.
[13]
Robert Geisberger, Michael N. Rice, Peter Sanders, and Vassilis J. Tsotras. 2012. Route planning with flexible edge restrictions. Journal
of Experimental Algorithmics 17, 1: 1.1. https://doi.org/10.1145/2133803.2133805
[14]
Reginald G. Golledge. 1995. Defining the criteria used in path selection. University of California Transportation Center.
[15]
Kevin Hamilton, Karrie Karahalios, Christian Sandvig, and Cedric Langbort. 2014. The image of the algorithmic city: a research
approach. Ann Arbor 1001: 48109.
[16]
Shazia Haque, Lars Kulik, and Alexander Klippel. 2006. Algorithms for reliable navigation and wayfinding. In International Conference
on Spatial Cognition, 308–326.
[17]
Brent Hecht, Johannes Schöning, Muki Haklay, Licia Capra, Afra J. Mashhadi, Loren Terveen, and Mei-Po Kwan. 2013. Geographic
human-computer interaction. In CHI ’13 Extended Abstracts on Human Factors in Computing Systems, 3163.
https://doi.org/10.1145/2468356.2479637
[18]
Steve Hendrix. 2016. Traffic-weary homeowners and Waze are at war, again. Guess who’s winning? Washington Post. Retrieved from
https://www.washingtonpost.com/local/traffic-weary-homeowners-and-waze-are-at-war-again-guess-whos-
winning/2016/06/05/c466df46-299d-11e6-b989-4e5479715b54_story.html?utm_term=.1ff02cceab51
[19]
Isaac Johnson, Connor McMahon, Johannes Schöning, and Brent Hecht. The Effect of Population and “Structural” Biases on Social
Media-based Algorithms–A Case Study in Geolocation Inference Across the Urban-Rural Spectrum.
[20]
David Jurgens, Tyler Finethy, James McCorriston, Yi Tian Xu, and Derek Ruths. 2015. Geolocation prediction in twitter using social
networks: A critical analysis and review of current practice. In ICWSM.
[21]
Jaewoo Kim, Meeyoung Cha, and Thomas Sandholm. 2014. SocRoutes: safe routes based on tweet sentiments. 179–182.
https://doi.org/10.1145/2567948.2577023
[22]
Chloe Kliman-Silver, Aniko Hannak, David Lazer, Christo Wilson, and Alan Mislove. 2015. Location, Location, Location: The Impact
of Geolocation on Web Search Personalization. 121–127. https://doi.org/10.1145/2815675.2815714
[23]
Julia Letchner, John Krumm, and Eric Horvitz. 2006. Trip router with individualized preferences (TRIP): Incorporating personalization
into route planning. In AAAI, 1795.
[24]
Shi-Chiang Li and Yongqiang Wu. 2012. South Florida Internet Based Route Choice Survey.
[25]
YiRu Li, Sarah George, Craig Apfelbeck, Abdeltawab M. Hendawi, David Hazel, Ankur Teredesai, and Mohamed Ali. 2014. Routing
Service with Real World Severe Weather. In SIGSPATIAL (SIGSPATIAL ’14), 585–588.
[26]
Ed Manley, Tao Cheng, and Andy Emmonds. 2011. Understanding Route Choice using Agent-based Simulation. Proceedings of the 11th
international conference on GeoComputation.
[27]
E.J. Manley, J.D. Addison, and T. Cheng. 2015. Shortest path or anchor-based route choice: a large-scale empirical analysis of minicab
routing in London. Journal of Transport Geography 43: 123–139. https://doi.org/10.1016/j.jtrangeo.2015.01.006
[28]
Aarian Marshall. 2016. Crime Alerts Come to Brazilian Waze, Just in Time for the Olympics. Wired. Retrieved from
https://www.wired.com/2016/07/crime-alerts-come-brazilian-waze-just-time-olympics/
[29]
Meghan McCarty. 2016. The Road Less Traveled? Not Since Waze Came To Los Angeles. NPR All Tech Considered. Retrieved from
http://www.npr.org/sections/alltechconsidered/2016/06/13/481871496/the-road-less-traveled-not-since-waze-came-to-los-angeles

15:20 • I. Johnson et al.
Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
[30]
Matt McFarland. 2014. The case for almost never turning left while driving. Washington Post. Retrieved from
https://www.washingtonpost.com/news/innovations/wp/2014/04/09/the-case-for-almost-never-turning-left-while-
driving/?utm_term=.9c07efa54881
[31]
Randy Meech. 2016. Open traffic data is the best traffic data. Retrieved from https://mapzen.com/blog/open-traffic-data/
[32]
Nikhil Naik, Jade Philipoom, Ramesh Raskar, and César Hidalgo. 2014. Streetscore-predicting the perceived safety of one million
streetscapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 779–785. Retrieved February
14, 2017 from http://www.cv-foundation.org/openaccess/content_cvpr_workshops_2014/W20/html/Naik_Streetscore_-
_Predicting_2014_CVPR_paper.html
[33]
Wassim G. Najm, John D. Smith, and David L. Smith. Analysis of Crossing Path Crashes. National Highway Traffic Safety
Administration.
[34]
Paul Newson and John Krumm. 2009. Hidden Markov map matching through noise and sparseness. In SIGSPATIAL, 336–343.
[35]
Ester Orban, Kelsey McDonald, Robynne Sutcliffe, Barbara Hoffmann, Kateryna B. Fuks, Nico Dragano, Anja Viehmann, Raimund
Erbel, Karl-Heinz Jöckel, Noreen Pundt, and Susanne Moebus. 2015. Residential Road Traffic Noise and High Depressive Symptoms
after Five Years of Follow-up: Results from the Heinz Nixdorf Recall Study. Environmental Health Perspectives 124, 5.
[36]
Christopher JD Patten, Albert Kircher, Joakim Östlund, Lena Nilsson, and Ola Svenson. 2006. Driver experience and cognitive workload
in different traffic environments. Accident Analysis & Prevention 38, 5: 887–894.
[37]
James W Pennebaker, Martha E Francis, and Roger J Booth. 2001. Linguistic inquiry and word count: LIWC 2001. Mahway: Lawrence
Erlbaum Associates 71, 2001.
[38]
Michal Piorkowski, Natasa Sarafijanovic-Djukic, and Matthias Grossglauser. 2009. CRAWDAD dataset epfl/mobility (v. 2009-02-24).
[39]
Linda Poon. 2016. Waze Puts Safety Over Speed by Minimizing Left Turns. CityLab. Retrieved from
http://www.citylab.com/tech/2016/06/waze-puts-safety-over-speed-by-minimizing-left-turns-in-la/487577/
[40]
Maaret Posti, Johannes Schöning, and Jonna Häkkilä. 2014. Unexpected journeys with the HOBBIT: the design and evaluation of an
asocial hiking app. 637–646. https://doi.org/10.1145/2598510.2598592
[41]
Daniele Quercia, Rossano Schifanella, and Luca Maria Aiello. 2014. The shortest path to happiness: recommending beautiful, quiet,
and happy routes in the city. In HT, 116–125.
[42]
Daniele Quercia, Rossano Schifanella, Luca Maria Aiello, and Kate McLean. 2015. Smelly maps: the digital life of urban smellscapes.
ICWSM.
[43]
Luca Rossi, Matthew J. Williams, Christoph Stich, and Mirco Musolesi. 2015. Privacy and the City: User Identification and Location
Semantics in Location-Based Social Networks. ICWSM.
[44]
Nina Runge, Pavel Samsonov, Donald Degraen, and Johannes Schöning. 2016. No more Autobahn!: Scenic Route Generation Using
Googles Street View. In IUI, 147–151.
[45]
Sumit Shah, Fenye Bao, Chang-Tien Lu, and Ing-Ray Chen. 2011. CROWDSAFE: Crowd Sourcing of Crime Incidents and Safe Routing
on Mobile Devices. In SIGSPATIAL (GIS ’11), 521–524.
[46]
Jie Shao, Lars Kulik, and Egemen Tanin. 2010. Easiest-to-reach neighbor search. In SIGSPATIAL, 360–369.
[47]
Jie Shao, Lars Kulik, Egemen Tanin, and Long Guo. 2014. Travel Distance Versus Navigation Complexity: A Study on Different Spatial
Queries on Road Networks. In Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge
Management (CIKM ’14), 1791–1794.
[48]
Monir H. Sharker, Hassan A. Karimi, and Janice C. Zgibor. 2012. Health-optimal Routing in Pedestrian Navigation Services. In
SIGSPATIAL (HealthGIS ’12), 1–10.
[49]
Gary Soeller, Karrie Karahalios, Christian Sandvig, and Christo Wilson. 2016. MapWatch: Detecting and Monitoring International
Border Personalization on Online Maps. 867–878. https://doi.org/10.1145/2872427.2883016
[50]
Hunter Stuart. 2016. Waze Lets Israelis Avoid Palestinian Areas, but Not the Other Way Around. Motherboard. Retrieved from
https://motherboard.vice.com/en_us/article/waze-lets-jewish-israelis-avoid-palestinian-areas-but-not-the-other-way-around
[51]
Ivan J Tashev, Jeffrey D Couckuyt, Neil W Black, John C Krumm, Ruston Panabaker, and Michael Lewis Seltzer. 2012. Pedestrian route
production. Google Patents.
[52]
Jérôme Thai, Nicolas Laurent-Brouty, and Alexandre M. Bayen. 2016. Negative externalities of GPS-enabled routing applications: A
game theoretical approach. In Intelligent Transportation Systems (ITSC), 2016 IEEE 19th International Conference on, 595–601.
[53]
Jacob Thebault-Spieker, Loren Terveen, and Brent Hecht. 2017. Towards a Geographic Understanding of the Sharing Economy:
Systemic Biases in UberX and TaskRabbit. ACM Transactions on Computer-Human Interaction.
[54]
Bart Thomee, David A. Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. 2016.
YFCC100M: The New Data in Multimedia Research. Commun. ACM 59, 2: 64–73. https://doi.org/10.1145/2812802
[55]
Martin Traunmueller, Ava Fatah gen. Schieck, Johannes Schöning, and Duncan P. Brumby. 2013. The Path is the Reward: Considering
Social Networks to Contribute to the Pleasure of Urban Strolling. In CHI EA (CHI EA ’13), 919–924.
https://doi.org/10.1145/2468356.2468520
[56]
Sudip Vhaduri, Amin Ali, Moushumi Sharmin, Karen Hovsepian, and Santosh Kumar. 2014. Estimating Drivers’ Stress from GPS
Traces. In Proceedings of the 6th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, 1–8.
[57]
Naomi Zeveloff. 2016. Israelis Sue Waze Navigation App for Creating Neighborhood Traffic Jam. Forward. Retrieved from
http://forward.com/news/israel/356487/israelis-sue-waze-navigation-app-for-creating-neighborhood-traffic-jam/
[58]
Jianwei Zhang, Hiroshi Kawasaki, and Yukiko Kawai. 2008. A Tourist Route Search System Based on Web Information and the
Visibility of Scenic Sights. In Proceedings of the 2008 Second International Symposium on Universal Communication (ISUC ’08), 154–161.
https://doi.org/10.1109/ISUC.2008.19
[59]
Yan-Tao Zheng, Shuicheng Yan, Zheng-Jun Zha, Yiqun Li, Xiangdong Zhou, Tat-Seng Chua, and Ramesh Jain. 2013. GPSView: A
Scenic Driving Route Planner. ACM Trans. Multimedia Comput. Commun. Appl. 9, 1: 3:1–3:18. https://doi.org/10.1145/2422956.2422959

Beautiful…but at What Cost? An Examination of Externalities in Geographic Vehicle Routing • 15:21
!
!
! Proceedings!of!the!ACM!on!Interactive,!Mobile,!Wearable,!and!Ubiquitous!Technologies,!Vol.!1,!No.!2,!Article!15.!Publication!date:!June!2017.!
[60]
Brian D. Ziebart, Andrew L. Maas, Anind K. Dey, and J. Andrew Bagnell. 2008. Navigate Like a Cabbie: Probabilistic Reasoning from
Observed Context-aware Behavior. In UbiComp (UbiComp ’08), 322–331. https://doi.org/10.1145/1409635.1409678
[61]
Dennis Zielstra and Hartwig Hochmair. 2012. Using free and proprietary data to compare shortest-path lengths for effective pedestrian
routing in street networks. Transportation Research Record: Journal of the Transportation Research Board, 2299: 41–47.
Received February 2017; revised April 2017; accepted June 2017