10 August 2022
AperTO - Archivio Istituzionale Open Access dell'Università di Torino
Original Citation:
The Shortest Path to Happiness: Recommending Beautiful, Quiet, and Happy Routes in the City
Publisher:
Published version:
DOI:10.1145/2631775.2631799
Terms of use:
Open Access
(Article begins on next page)
Anyone can freely access the full text of works made available as "Open Access". Works made available
under a Creative Commons license can be used according to the terms and conditions of said license. Use
of all other works requires consent of the right holder (author or publisher) if not exempted from copyright
protection by the applicable law.
Availability:
ACM - Association for Computing Machinery
This is a pre print version of the following article:
This version is available http://hdl.handle.net/2318/155303 since
This is an author version of the contribution published on:
Daniele Quercia, Rossano Schifanella, Luca Maria Aiello
The Shortest Path to Happiness: Recommending Beautiful, Quiet, and Happy
Routes in the City
Editor: ACM - Association for Computing Machinery
2014
ISBN: 9781450329545
in
Proceedings of the 25th ACM conference on Hypertext and social media - HT
'14
116 - 125
25th ACM Conference on Hypertext and Social Media
Santiago, Chile
September 1-4
The definitive version is available at:
http://dl.acm.org/citation.cfm?doid=2631775.2631799
The Shortest Path to Happiness:
Recommending Beautiful, Quiet, and Happy Routes in the
City
Daniele Quercia
Yahoo Labs
Barcelona, Spain
dquercia@yahoo-inc.com
Rossano Schifanella
∗
University of Torino
Torino, Italy
schifane@di.unito.it
Luca Maria Aiello
Yahoo Labs
Barcelona, Spain
alucca@yahoo-inc.com
ABSTRACT
When providing directions to a place, web and mobile map-
ping services are all able to suggest the shortest route. The
goal of this work is to automatically suggest routes that are
not only short but also emotionally pleasant. To quantify
the extent to which urban locations are pleasant, we use data
from a crowd-sourcing platform that shows two street scenes
in London (out of hundreds), and a user votes on which one
looks more beautiful, quiet, and happy. We consider votes
from more than 3.3K individuals and translate them into
quantitative measures of location perceptions. We arrange
those locations into a graph upon which we learn pleasant
routes. Based on a quantitative validation, we find that,
compared to the shortest routes, the recommended ones add
just a few extra walking minutes and are indeed perceived
to be more beautiful, quiet, and happy. To test the gener-
ality of our approach, we consider Flickr metadata of more
than 3.7M pictures in London and 1.3M in Boston, com-
pute proxies for the crowdsourced beauty dimension (the
one for which we have collected the most votes), and evalu-
ate those proxies with 30 participants in London and 54 in
Boston. These participants have not only rated our recom-
mendations but have also carefully motivated their choices,
providing insights for future work.
Categories and Subject Descriptors
H.4 [Information Systems Applications]: Miscellaneous
General Terms
Human Factors, Design, Measurement.
∗
This work has been done while the author was visiting Ya-
hoo Labs, Barcelona, within the framework of the Faculty
Research and Engagement Program.
Permission to make digital or hard copies of all or part 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 components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
HT’14, September 1–4, 2014, Santiago, Chile.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-2954-5/14/09 ...$15.00.
http://dx.doi.org/10.1145/2631775.2631799 .
Keywords
Social Media; Urban Informatics; Derives
1. INTRODUCTION
At times, we do not take the fastest route but enjoy alter-
natives that offer beautiful sceneries. When walking, we gen-
erally prefer tiny streets with trees over large avenues with
cars. However, Web and mobile mapping services currently
fail to offer that experience as they are able to recommend
only shortest routes.
To capture which routes people find interesting and enjoy-
able, researchers have started to analyze the digital traces
left behind by users of online services like Flickr or Four-
square. Previous work has, however, not considered the
role of emotions in the urban context when recommending
routes. Yet, there exists the concept of psychogeography,
which dates back to 1955. This was defined as “the study of
the precise laws and specific effects of the geographical envi-
ronment, consciously organized or not, on the emotions and
behavior of individuals” [9]. The psychogeographer “is able
both to identify and to distill the varied ambiances of the
urban environment. Emotional zones that cannot be deter-
mined simply by architectural or economic conditions must
be determined by following the aimless stroll (derive)” [6].
Mobile applications have been recently proposed to ease
making derives (i.e., detours in the city): these include De-
rive app
1
, Serendipitor
2
, Drift
3
, and Random GPS.
4
The
goal of our work is to go beyond supporting derives and to
propose ways of automatically generating routes that are not
only short but also emotionally pleasant. This goal is not
algorithmic but is experimental. We rely on crowdsourced
measurements of people’s emotional experience of the city
and use those measurements to propose new ways of recom-
mending urban routes. Despite emotional responses being
subjective and difficult to quantify, urban studies have re-
peatedly shown that specific visual cues in the city context
are consistently associated with the same fuzzy concept (e.g.,
with beauty) [7, 23, 25, 28]. For example, previous work has
found that green spaces and Victorian houses are mostly as-
sociated with beauty, while trash and broken windows with
ugliness.
1
http://deriveapp.com/s/v2/
2
http://serendipitor.net/site/
3
http://www.brokencitylab.org/drift/
4
http://nogovoyages.com/random_gps.html
arXiv:1407.1031v1 [cs.SI] 3 Jul 2014
To meet our research goal, we make three main contribu-
tions:
• We build a graph whose nodes are locations and whose
edges connect geographic neighbors (§3.1). With this
graph, we rank locations based on whether they are
emotionally pleasant. The emotion scores come from
a crowd-sourcing platform that shows two street scenes
in London (out of hundreds), and a user votes on which
one looks more beautiful, quiet, and happy (§3.2).
• We quantitatively validate the extent to which our pro-
posal recommends paths that are not only short but
also emotionally-pleasing (§4). We then qualitatively
evaluate the recommendations by conducting a user
study involving 30 participants in London.
• We finally test the generalizability of our proposal by:
a) presenting a way of predicting the beauty scores
from Flickr metadata; and b) testing the beauty-deri-
ved paths with our 30 participants in London and with
a new group of 54 participants in Boston (§5).
2. RELATED WORK
Early research on route recommendation focused on find-
ing the most efficient routes. For example, Chang et al. ([2])
used a backtracking algorithm to recommend car routes that
deviate from a user’s familiar/past trajectories. Ludwig et
al. ([16]) used an adaptive A*-like algorithm to recommend
public transport routes that afford both short walks and lit-
tle waiting times. More recently, tools for recommending the
safest or smoothest cycle paths in the city have also been
proposed [24].
In addition to ways of recommending efficient paths, re-
searchers have also investigated the problem of recommen-
ding distinctive and interesting urban routes [22, 31]. The
idea behind this line of work is to use geo-referenced online
content (e.g., Flickr pictures) to learn and recommend pop-
ular trajectories [1, 34, 35]. De Choundry et al. ([8]) and El
Ali et al. ([10]) both used Flickr data to mine popular spatio-
temporal sequences of picture uploads and to then recom-
mend the corresponding urban routes. De Choundry et al.
identified the movements of individual tourists by tracking
when and where they were uploading photos and by then
using the resulting trajectories to connect points of inter-
ests in a graph. By embedding location information such
as average time spent at a location and location popularity,
they were able to use an orienteering algorithm on the graph
to compute the optimal number of interesting locations to
visit given a time budget. El Ali et al. followed a similar
idea: by clustering sequences of pictures uploaded at similar
times, they used a sequence alignment algorithm borrowed
from biology to produce trajectories containing interesting
locations.
In addition to using geo-located pictures, one could exploit
GPS traces. As opposed to social media, mobile phones en-
joy high penetration rates and, as such, GPS traces can help
identify interesting places not only in cities but also in sub-
urban regions [32]. Zheng et al. ([36]) mined such traces by
arranging both visited places and mobile users in a bipar-
tite graph, and then ranking places by graph centrality to
extract top interesting regions. Their focus was on spotting
interesting locations rather than routes.
All this line of work has been tailored to touristic use cases
where paths can be considerably longer than the shortest
ones [29], and where recommending frequently visited loca-
tions is a reasonable choice.
More recently, given the popularity of mobile social net-
working applications, researchers have been able to explore
personalization strategies for tourists and residents alike.
Meng et al. ([18]) leveraged traces from Foursquare to plan
itineraries that need to pass through different types of lo-
cations (e.g., restaurants, gas stations) and, given a user
demand for some location types, they computed paths using
an ant colony optimization algorithm. Cheng et al. ([3]) an-
notated historical data of traveled paths with demographic
information and used a Baeysian learning model to generate
personalized travel recommendations based on demographic
segmentation. Kurashima et al. ([15]) addressed a somewhat
similar problem - they profiled users according to their past
travel histories. These approaches output sequences of loca-
tions according to different criteria but do not focus on the
nature of the paths connecting those locations.
To date, there has not been any work that considers peo-
ple’s emotional perceptions of urban spaces when recommen-
ding routes to them. We thus set out to do such a work by
collecting reliable perceptions of urban scenes, incorporat-
ing them into algorithmic solutions, and quantitatively and
qualitatively evaluating those solutions.
3. OUR PROPOSAL
Our goal is to suggest users a short and pleasant path
between their current location s and destination d. We meet
this goal in four steps:
1. Build a graph whose nodes are all locations in the city
under study (§3.1).
2. Crowdsource people’s perceptions of those locations
along three dimensions: beautiful, quiet, and happy
(§3.2).
3. Assign scores to locations along each of the three (§3.3).
4. Select the path between nodes s and d that strikes the
right balance between being short and being pleasant
(§3.4).
3.1 Building Location Graph
We divide the bounding box of central London (travel zone
1
5
) into 532 walkable cells, each of which is 200x200 meters
in size. Previous research has established that 200m tends to
be the threshold of walkable distance in urban areas [20, 4].
In dense parts of London, such a distance would typically
correspond to two blocks that could be covered by a 2.5-
minute walk. Having those cells at hand, we make them be
nodes in a location graph. Each node is a location and links
to its eight geographic neighbors.
6
To quantify the extent to
which a node reflects a pleasant location, we need to capture
the way people perceive that location, and we do so next.
3.2 Crowdsourcing Perceptions
We rely on the data gathered from a crowdsourcing web
site to assess the extent to which different city’s locations are
5
http://visitorshop.tfl.gov.uk/help-centre/about-travel- zones.html
6
Since a boundary cell would have less than eight neighbors, we link it to a
number of additional closest cells within the grid such that, as a result, it would
link to eight nodes in total.
perceived to be beautiful, quiet and make people happy [25].
Available under UrbanGems.org, the site picks up two ran-
dom urban scenes and ask users which one of the two is
more beautiful, quiet, or happy. As for scenes, the site does
not use Flickr images, as they considerably vary in quality,
but taps into two image sources that offer pictures of com-
parable quality: Google Street View pictures captured by
camera-mounted cars, and Geograph
7
pictures provided by
volunteers. To control for image bias, we perform two main
steps. First, we make sure that multiple images from the
two sources are available at each location. Second, we check
whether user ratings are not correlated with objective mea-
sures of image quality, and we indeed find that there is no
correlation between images’ ratings and two commonly used
proxies for quality (i.e., sharpness and contrast levels [33]).
At each game round, users should either click on one of
the two scenes or opt for “Can’t Tell”, if undecided on which
picture to click on. With each selection, the user is asked
to guess the percentage of other people who shared their
views, scoring points for correct guesses. Those points can
then be shared through the social media sites of Facebook
and Twitter. To avoid the sparsity problem (too few answers
per picture), a random scene is selected within a 300-meter
radius from a subway station and within the bounding boxes
of census areas. This results into 258 Google Street views
and 310 Geograph images, all of which have ratings that are
roughly normally distributed [25]. We use multiple images
from the two sources at each location. By collecting a large
number of responses across a large number of participants,
we are now able to determine which urban scenes are per-
ceived in which ways along the three qualities.
The choice of the three qualities is motivated by their im-
portance in the urban context according to previous studies.
Being able to find quiet places might “promote ‘sonic health’
in our cities and offer a public guide for those who crave a
retreat from crowds”
8
. As for beauty, we are not the first
to measure its perceptions. In 1967, Peterson proposed a
quantitative analysis of public perceptions of neighborhood
visual appearance [23] and found that beauty and safety are
approximately collinear. Finally, we choose happiness not
least because urban studies in the 1960s tried to systemati-
cally relate well-being in the urban environment (i.e., happi-
ness) to the fundamental desire for visual order, beauty, and
aesthetics [17]. As a result, well-being or, more informally,
happiness has taken centre stage in the scientific discourse
for decades.
The platform was released in September 2012 and after
4 months data from as many as 3,301 participants was col-
lected: 36% connecting from London (IP addresses), 35%
from the rest of UK, and 29% outside UK. A fraction of
those participants (515) specified their personal details: the
percentage of male-female for those participants is 66%-34%,
the average age is 38.1 years old, and the racial segmenta-
tion reflects that of the 2001 UK census.
9
Upon processing
17,261 rounds of annotation (each round requires to anno-
tate at most ten pairs), we rank pictures by their scores for
beauty, quiet, and happiness, and those scores are based on
the fraction of votes the pictures have received. The rank-
ing is reliable because the number of annotators is >3K and
distribution of scores is normal with median as high as 171
7
http://www.geograph.org.uk/
8
http://www.stereopublic.net/
9
http://en.wikipedia.org/wiki/Ethnic_groups_in_the_United_Kingdom
Name Formula
Linear h
i
Cubic h
3
i
Exponential e
h
i
Square root
√
h
i
Sigmoid
1
1+e
−h
i
Table 1: Five expressions experimentally used to
map crowdsourced scores to probabilities. With a
location’s crowdsourced scores of happiness, those
expressions return the location’s likelihoods of being
considered happy if one were to visit it.
for beauty, 12 for quiet, and 16 for happy. The number
of answers for the three qualities is different as the default
question is that on beauty, which thus preferentially attracts
more answers.
We compute the correlations between each pairwise com-
binations of the three qualities. All correlations are statisti-
cally significant (i.e., all p-values are < 0.0001) and are the
following: happy-quiet has r = 0.29, quiet-beauty r = 0.33,
and beauty-happy is r = 0.64. As one would expect from
the literature [7], we find that the strongest affiliation is that
between beauty and happiness, so we should expect that the
paths we will recommend for beauty and those for happiness
might partly overlap at times.
3.3 Scoring Locations
To rank a location, we need to compute the likelihood that
it will be visited because it is pleasant. One simple way of
expressing that is with p(go|happy) ∝ p(happiness|go).
The probability p(happiness|go) captures the idea that
individuals visits locations that make them happy. We thus
need a way to measure a location’s happiness and, to that
end, we resort to our crowdsourced scores (§3.2). More
specifically, given the crowdsourced happiness score h
i
for
cell i, we compute the corresponding happiness probability
with a curve that is, for example, cubic:
p(happiness|go) = k · h
3
i
, where k =
1
max{h
3
i
}∀
i
(1)
Thus, the higher a location’s crowdsourced score, the hap-
pier it is likely to be. By substituting h
3
i
with any of the
expressions in Table 1, we obtain the alternative happiness
scoring functions. We apply those scoring functions to the
two remaining scores of quiet and beauty too. We do so by
simply substituting h
i
with q
i
(location i’s quietness score)
and with b
i
(i’s beauty score). Next, for brevity, we will
report only the results for the cubic curve for which our ex-
periments have shown the highest percentage improvements.
3.4 Selecting Best Path
Upon the location graph and having the likelihood of vis-
iting each location, we now select the best path from source
s to destination d in four steps:
Step 1. Identify M shortest paths between s and d. To
identify them, we run Eppstein’s algorithm [11] and find the
M shortest paths connecting each pair of nodes s and d. To
be sufficiently “exhaustive”, we initially set M to be as high
as 10
6
. This choice makes it possible to explore the full set
