Content uploaded by Iván Cantador
Author content
All content in this area was uploaded by Iván Cantador on Feb 21, 2018
Content may be subject to copyright.
Content uploaded by Iván Cantador
Author content
All content in this area was uploaded by Iván Cantador on Sep 26, 2017
Content may be subject to copyright.
Content uploaded by María E. Cortés-Cediel
Author content
All content in this area was uploaded by María E. Cortés-Cediel on Aug 10, 2017
Content may be subject to copyright.
Content uploaded by Olga Gil
Author content
All content in this area was uploaded by Olga Gil on Sep 07, 2017
Content may be subject to copyright.
Personalized recommendations in e-participation:
Offline experiments for the ‘Decide Madrid’ platform
Iván Cantador1, Alejandro Bellogín1, María E. Cortés-Cediel2, Olga Gil2
1
Escuela Politécnica Superior
Universidad Autónoma de Madrid
Spain
{ivan.cantador, alejandro.bellogin}@uam.es
2
Facultad de Ciencias Políticas y Sociología
Universidad Complutense de Madrid
Spain
{mcorte04, olgagil}@ucm.es
ABSTRACT
In e-participation platforms, citizens suggest, discuss and vote
online for initiatives aimed to address a wide range of issues and
problems in a city, such as economic development, public safety,
budges, infrastructure, housing, environment, social rights, and
health care. For a particular citizen, the number of proposals and
debates may be overwhelming, and recommender systems could
help filtering and ranking those that are more relevant. Focusing
on a particular case, the ‘Decide Madrid’ platform, in this paper
we empirically investigate which sources of user preferences and
recommendation approaches could be more effective, in terms of
several aspects, namely precision, coverage and diversity.
KEYWORDS
recommender systems, e-participation, citizen participation,
urban planning, smart cities, social tagging, folksonomy
1 INTRODUCTION
1
Plans related to smart cities are drafted to mitigate and remedy
urban challenges and problems in a sustainable way through
innovation [2]. They commonly entail a strong integration of
information and communications technologies (ICTs) into
planning, operations and management. In addition, interaction
and participation of citizens and residents are common when
digital media are set in place.
In this context, e-participation platforms are commonly used in
smart cities in order to upgrade the relations among stakeholders
in civil society –including citizens, residents, firms and the local
government itself–, and perform as a mechanism to put the
citizens at the center of the process [1]. This all reflects that at the
core of the smart city there is an attempt to develop new forms of
collaboration and urban development through ICTs [17].
One of the pillars of smart city plans is governance. Local
governance within the smart cities encapsulates collaboration,
1 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.
2
3 CitRec, August 27, 2017, Como, Italy
4 © 2017 Copyright is held by the owner/author(s). Publication rights licensed to
ACM.
5 ACM ISBN 978-1-4503-5370-0/17/08...$15.00
6 https://doi.org/10.1145/3127325.3127330
cooperation, partnerships and participation, which might
become success factors in the city [18]. Hence, when we talk
about governance in the smart city, we refer to the adoption of
inclusive and participatory processes that allow for the
deliberation of social actors [12], and their implication in the
formulation of public policies. Governance thus involves
multiple interactions between the different stakeholders, public
institutions, citizens, researcher institutes, and firms.
Adding ICTs to governance might facilitate better exchange of
information and sustainability in such interactions [9]. In
particular, the modes of participatory interactions involving local
governments with citizens and residents might be characterized
as three-fold. In each mode, ICTs are applied and used to a
different extent. On the first mode, interactions reach the level of
information; on a second mode, interactions would involve
e-consultation; and on the third mode, interactions would entail
e-participation, which requires the greater degree of involvement.
Moreover, according to the scientific literature, citizenship
participation in public affairs is related to three relevant aspects.
First of all, the improvement of democratic legitimacy in
increasingly complex societies. Secondly, the improvement of
effectiveness and efficiency of public policies and, lastly, the
development of an active citizenship through experiences of
participation [4].
Reports produced by the OECD [20][21] have covered the issue
of increasing citizen participation in politics through ICTs. The
theoretical frameworks in these reports focus on four objectives
[15]: (i) promoting participation through a widened audience; (ii)
seeking participation from citizens and residents through ICTs to
leap forward on technical and communication skills; (iii)
facilitating relevant information through a more accessible
format for audiences; and (iv) engaging in deliberative process
with an ample majority. These objectives have been assumed –to
a different extent– by governments using ICTs as a way to
increase public participation and the possibility to enhance the
benefits for citizens [11].
In so doing, there are convenient strategies to identify needs of
the citizenship, and to provide the tools for participation [23].
E-participation platforms are institutionalized mechanisms that
allow the citizens to participate in democratic life, and thus be an
active part of government plans and decisions [7]. However,
these platforms have a number of problems, such as an excess of
information, and a requirement for customization.
For a particular citizen, the number of initiatives and discussions
in an e-participation platform may be overwhelming. Addressing
this situation, recommender systems could filter and rank the
initiatives and discussions that are more relevant for the citizens
based on previous explicit interests and analyzed implicit
CitRec’17, August 2017, Como, Italy
I. Cantador, A. Bellogín, M. E. Cortés-Cediel, O. Gil
behavior. In this way, they may not only promote the citizens’
participation, but also could increase their engagement.
This work intends to proof the usefulness of personalized
recommendations in e-participation. It does so through offline
experiments carried out over the ‘Decide Madrid’ platform
2
, which
is the online digital medium set in place by the local government
in the city of Madrid. Aiming to investigate which sources of user
preferences (e.g., comments and social tags) and recommendation
approaches (e.g., content-based and collaborative filtering) could
be more effective, we perform offline experiments on a dataset
obtained from the ‘Decide Madrid’ platform, and with a variety of
evaluation metrics, such as precision, coverage and diversity.
2 RELATED WORK
Within the different forms of e-governance, government-to-
citizens (G2C) governance aims to provide citizens with a variety
of online information and e-services in an efficient and cost-
effective manner, and to strengthen the relationship between
government and citizens with ICTs. This is addressed at different
levels of interaction between such actors, distinguishing among
information, consultation and participation levels [20].
At the e-information level, the government offers websites with
information on policies and programs, laws and regulations,
budgets, and other issues of public interest, as well as software
tools –such as email subscription lists, online newsgroups, and
web forums– for the dissemination, and timely access and use of
public information and services. In this context, recommender
systems have been mainly proposed to provide the citizens with
personalized government e-notifications and e-services; see e.g.
[3], [5] and [6].
At the e-consultation level, the government offers online
consultation (a.k.a. e-voting) mechanisms and tools, which present
citizens with choices about public policy topics, allowing for the
deliberation in real time, as well as the access to archived audios
and videos of public meetings. Citizens are thus encouraged to
contribute to the government consultations. In this context,
recommender systems could assist voters in making decisions by
providing recommendations about candidates close to the voters’
preferences and tendencies. Terán and Meier [22] proposed a
recommendation framework aimed to assist voters in making
decisions by providing information about candidates close to the
voters’ preferences and tendencies. Its recommendations are based
on similarities between voters and candidates –whose profiles are
created by filling a questionnaire about values, attitudes and
political issues on a number of topic categories. The system
performs a fuzzy based clustering algorithm, and generates a
graphical representation of political parties distributed in
generated clusters, helping citizens to analyze politicians.
Finally, at the e-participation level, the government provides
online participation platforms where citizens can propose,
discuss, give feedback, and vote for initiatives aimed to solve or
improve a wide range of situations and problems in different
aspects of a city, such as health and social care, culture and
education, energy and environment, and urban mobility and
transport. In these platforms, recommendation approaches can
assist the citizens in finding relevant proposals, discussions,
individuals and associations, according to personal interests
explicitly declared though votes, or implicitly expressed by means
2
‘Decide Madrid’ e-participation platform, https://decide.madrid.es
of online comments and social links. Nelimarkka et al. [19]
present CRC, an online civic engagement platform that,
differently to other analogous platforms, facilitates the
participants’ consideration of diverse viewpoints, an issue that is
desirable in democratic processes and increases civic engagement,
as shown by the authors. For such purpose, the platform
recommends comments from individuals who hold similar and
dissimilar opinions, by means of uncertainty minimizing
sampling and PCA techniques.
Kavanaugh et al. [14] present Virtual Town Square, a location-
based information aggregator system aimed to support and
facilitate citizens’ discussion and interaction. The system
captures such information from several local news providers and
user-generated media. Then, it filters and recommends relevant
information items according to several aspects, such as topic,
social media popularity, citizens’ comments, and collaborative
filtering similarities with like-minded people within trusted
groups. Finally, instead of providing recommendation to citizens,
Marsal-Llacuna and De la Rosa-Esteva [16] propose an agent-
based model that, mining citizens’ opinions expressed on the web,
makes recommendations to planners on the design of an urban
plan. A particular innovative feature of the model is that public
participation occurs before and during the design and
development of the plan. The recommendations are generated by
a demographic collaborative filtering agent that exploits citizens’
satisfaction surveys concerning a variety of issues about the city,
and a content-based filtering agent that mines opinions of citizens
from others cities about projects related to the target plan.
3 THE ‘DECIDE MADRID’ PLATFORM
In September 2015, Madrid city council launched the ‘Decide
Madrid’ e-participation platform, a web system designed to allow
Madrid residents to make, debate and vote proposals for the city
on a variety of topics, such as transport, natural environment,
urbanism, social rights, health care, education, and culture.
Through this system, citizens contribute and decide how to
spend part of the assigned participatory budget, which has been
set to 100 million Euros for 2017.
This process consists of three main phases, namely submit, support
and vote phases. In the submit phase, any person can create a
proposal by signing up on the platform, and filling a simple
questionnaire specifying a title, description, and some optional
tags for the proposal (Figure 1 left). Then, the support phase is
aimed to prioritize the most interesting and relevant proposals. For
such purpose, city residents who are over 16 years old are allowed
to explicitly express their support to existing proposals. The
proposals that get support from 27064 people (i.e., 1% of the
allowed residents) in a period of 30 days are approved. Before
passing to the next phase, during a period of 45 days, approved
proposals are commented and discussed by the citizens in the
platform (Figure 1 right). The proposals without enough support
are discarded and archived. Finally, in the vote phase, during a
period of one week from its approval date, each approved proposal
can be voted by allowed residents. In case there are more people in
favor than against, a proposal is accepted as a ‘collective proposal’
of Madrid citizens, and the city council government assumes it as
its own and carries it out. To achieve this, within a maximum
period of one month, the corresponding technical reports on the
legality, feasibility and economic cost of the proposal are
published on the web. Then, citizens can access the plan to
accomplish the proposal, and track its progress.
Personalized recommendations in e-participation: Offline experiments for the Decide Madrid platform CitRec’17, August 2017, Como, Italy
Figure 1: Screenshots of the ‘Decide Madrid’ platform, showing proposals metadata (title, author, date, description, tags,
number of supports) and citizen comments.
3.1 The ‘Decide Madrid’ folksonomy
When a citizen creates a proposal in the ‘Decide Madrid’
e-participation platform, she has the opportunity to annotate it
with a number of freely chosen words, which are called as
categories in the platform, and are commonly known as social
tags in the literature.
In a particular system, the whole set of tags constitutes an
unstructured collaborative classification scheme that is referred
as folksonomy. This implicit classification is then used to search,
discover and recommend (tagged) resources of interest.
In general, within the ‘Decide Madrid’ platform, the tags
assigned by a user to a particular proposal correspond to the
places and topics related with the proposal. In both cases, a tag,
which is a plain-text word, does not follow any categorization
schema, and has not been assigned any metadata, which would
allow establishing the meaning, type and properties of the
concept underlying the tag. Hence, a ‘place tag’ may refer to a
district (e.g., centro), a neighborhood (e.g., sol), a street (e.g., gran
via), a square (e.g., plaza mayor), and a museum (e.g., el prado),
to name a few; and a ‘topic tag’ may refer to a city asset (e.g.,
transporte), a particular issue (e.g., precio metro), and a social
group (e.g., jovenes), among others.
As we shall explain in the next section, we have developed
methods to determine whether a tag refers to a place or to a
topic, and specify its corresponding concept and type. In our
experiments, we evaluate content-based and hybrid
recommendation approaches that exploit tags with and without
the above processing.
3.2 The ‘Decide Madrid’ proposal forums
As mentioned before, and it is shown in Figure 1 (right), the
‘Decide Madrid’ platform provides an online forum for each
proposal where citizens discuss it by means of comments and
replies threads. In the forum, every comment and reply may
receive positive and negative votes from the system users.
The platform website shows the number of positive/negative
votes given to all proposals, comments and replies. In a
collaborative filtering context, the votes would be the citizens’
ratings for the proposals.
However, the website does not show the users who gave such
votes, and thus does not make the explicit [user, item, rating]
tuples publicly available. To address this situation, in our
experiments, we consider the users’ comments as a signal of
their interest for the corresponding proposals, and thus we treat
them as ratings, for collaborative filtering offline evaluation.
4 EXPERIMENTS
4.1 Dataset
We conducted the experiments on a dataset generated with data
publicly available in the ‘Decide Madrid’ website. Specifically,
we crawled the website gathering information about the
proposals recorded in the system from 15th September 2015 to
31st May 2017. To obtain the users’ comments, we only accessed
the first page of each proposal forum, according to the
decreasing popularity of its comments.
Table 1 shows statistics about the dataset. We obtained a total of
54357 ratings (i.e., unique [user, proposal] comment pairs) given
by 17991 users to 16880 proposals, which leads to a 99.98% of
rating sparsity. All the proposals were tagged. The users provided
58294 tag assignments (i.e., unique [user, tag, proposal] tuples),
using 2967 distinct tags, from which the 7.65% were mapped to
places, and 52.28% were mapped to topics. In the next two
subsections, we explain how we performed such tag mappings.
Number of users
17991
Number of users who created proposals
11489
Number of users who commented proposals
10481
Number of proposals
16880
Number of proposals with place tags
16880
(100%)
Number of proposals with topic tags
11724
(69.45%)
Number of tags
2967
Number of place tags
227
(7.65%)
Number of topic tags
1551
(52.28%)
Number of tag assignments
58294
Number of place tag assignments
24179
(41.48%)
Number of topic tag assignments
31691
(54.36%)
Number of ratings
54357
Number of proposals with ratings
12055
(71.42%)
Rating sparsity
99.98%
Table 1: Dataset statistics.
CitRec’17, August 2017, Como, Italy
I. Cantador, A. Bellogín, M. E. Cortés-Cediel, O. Gil
4.1.1 Place tags
To determine whether each plain-text tag corresponds to a
certain place in Madrid, we first created a repository of places in
the city. Madrid has 21 districts (Figure 2), each of them with
several neighborhoods. In total, it has 129 neighborhoods. We
also considered an artificial district we called ‘whole city’ since
there are proposals that are applicable to all the districts.
From a public database available at the ‘madrid.org’ Open Data
portal
3
, we downloaded lists of streets (and their neighborhoods)
and representative places (e.g., hospitals and museums) in Madrid.
Then, we processed the database transforming the names of the
places to words in lowercase without numbers, punctuation
symbols and accented vowels (as we also did with the social tags
from the platform), e.g., Gran Vía, 1 was transformed to gran via.
Table 2 shows statistics about the generated places database.
Type
Places
Type
Places
Districts
22
Police stations
50
Neighborhoods
129
Shopping centers
48
Streets, squares, bridges
7979
Cinemas
36
Universities, colleges, schools
543
Parks
31
Religious buildings
317
Cemeteries
22
Hospitals, community health centers
170
Fire stations
12
Theaters
113
Amusement parks
4
Museums
63
Bullrings
1
Table 2: Places database statistics.
Afterwards, we mapped tags (and consequently proposals) to
places by exact matching with the places names. As can be
observed in Table 3, there was a quite uniform distribution of the
citizens’ proposals along the city districts.
Figure 2: Districts of Madrid. The numbers are district ids.
Id
Name
Proposals
Id
Name
Proposals
0
Whole city
13118
(77.71%)
13
Puente de Vallecas
184
(1.09%)
1
Centro
377
(2.23%)
11
Carabanchel
181
(1.07%)
17
Villaverde
353
(2.09%)
19
Vicálvaro
166
(0.98%)
8
Fuencarral-El Pardo
351
(2.08%)
3
Retiro
147
(0.87%)
18
Villa de Vallecas
311
(1.84%)
15
Ciudad Lineal
141
(0.84%)
16
Hortaleza
286
(1.69%)
6
Tetuán
125
(0.74%)
10
Latina
246
(1.46%)
4
Salamanca
113
(0.67%)
12
Usera
230
(1.36%)
14
Moratalaz
101
(0.60%)
2
Arganzuela
219
(1.30%)
7
Chamberí
98
(0.58%)
9
Moncloa-Aravaca
205
(1.21%)
5
Chamartín
88
(0.52%)
20
San Blas-Canillejas
202
(1.20%)
21
Barajas
85
(0.50%)
Table 3: Number of proposals per district.
3
Madrid places, http://www.madrid.org/nomecalles/DescargaBDTCorte.icm
4.1.2 Topic tags
The ‘Decide Madrid’ platform has 16 tags as main categories for
the proposals. However, many other categories could be
considered as topics of interest for the citizens, and several tags
may refer to the same category. For these reasons, we extended
the number of topic categories, and performed a semi-automatic
method for assigning existing tags to each category.
Specifically, we manually inspected the 150 most popular tags in
our dataset, and grouped them in 30 tag sets, each of them
representing a topic of interest. The selection of the 30 topics
was done during the above inspection.
Then, iteratively and following a decreasing popularity ordering,
for every remaining tag t, we computed its Levenshtein distance
with the tags assigned to each category. Next, the tag t was
assigned to certain category if the category contained (i) a tag
with distance lower than 3 to t –e.g., alcalde and alcaldesa for the
‘City hall & Public Administration’ category–, or (ii) a tag that
contains t as substring, e.g., the tag accesibilidad metro was
assigned to ‘Accessibility’ and ‘Transport’ categories, since
accesibilidad and metro already belonged to such categories,
respectively.
For each category, Table 4 shows the number of tags assigned to
the category, and the number of proposals tagged with at least
one tag of the category. Issues in transport and urban mobility
(e.g., traffic jams, and public parking), natural environment and
sustainability (e.g., pollution and waste), and health care (e.g.,
hospital resources) are well known, major problems in Madrid.
Topic category
Proposals
Tags
Topic category
Proposals
Tags
Transport
4372
237
City hall, Public Administration
444
70
Natural environment
4092
210
Education
268
94
Urbanism
2932
139
Animals
168
98
Health care
2246
125
Family and childhood
133
48
Sustainability
2187
39
Civic virtue
131
50
Social rights
1974
99
Justice
112
46
Citizen participation
1825
38
Leisure, entertainment
105
46
Culture
1638
53
Accessibility
83
23
Economy
1324
66
Politics
66
43
Sports
1210
73
Housing
60
31
Security, emergencies
1077
57
Adolescence, youth
56
21
Equity and integration
1025
68
Tourism
62
8
Government transparency
932
18
Delinquency
49
34
Job
886
30
Old age
33
22
Associations
603
21
Religion
11
6
Table 4: Number of tags per topic.
4.2 Citizen and proposal profiles
As usually done in content-based recommender systems, we
defined user (citizen) and item (proposal) profiles as vectors in
the same space. In particular, we built the profiles according to
the tags used by the users to annotate the items.
We experimented with both binary and weighted tag profiles for
the items, and the binary versions were consistently worse, in
agreement with previous work [10]. As we will show in Section
4.5, we tested TF-IDF and BM25 weighting techniques, as done in
[10]. Furthermore, since there are tags that are intrinsically
related to the proposals by indicating specific topics and affected
places, we also experimented with item profiles containing one
specific tag type, i.e., place tags or topic tags. Note that this
Personalized recommendations in e-participation: Offline experiments for the Decide Madrid platform
CitRec’17, August 2017, Como, Italy
filtering shrinks the vector space, favoring less sparse
representations of the items.
Once the item profiles were built, we generated the user profiles
by considering the proposals ‘rated’ by each citizen, since they
describe her interests and tastes. Specifically, for each user, we
aggregated her rated item profiles, and accumulated the weights
computed for each dimension. By performing this
transformation, we obtain a vector representation of users in the
same space as items, allowing comparisons between them.
4.3 Recommendation algorithms
We evaluated the following recommendation algorithms,
implemented on top of the RankSys framework
4
:
cb: a content-based (CB) recommendation approach that
exploits social tags to build the user and item profiles (as
explained in Section 4.2). The score produced by this method
is the Cosine similarity between the user’s profile and the
profile of every proposal (not previously seen by the user) in
the system. As described before, in addition to exploiting all
tags, we have performed different tag mappings, which are
transparently evaluated by generating different user and
item profiles, where, for instance, place vs. topic tags are
compared.
ub: a user-based nearest neighbor approach (UB k-NN) that
exploits the rating-based similarity between users to create
neighborhoods, which are used to compute a score for each
(user, item) pair. In the experiments, we used the Cosine
similarity between users and several neighborhood sizes,
namely k = 5, …, 100, in steps of 5.
mf: a matrix factorization collaborative filtering algorithm.
We decided to use the variation proposed in [13] (the HKV
factorizer implemented in RankSys,), since it is well suited
for implicit feedback datasets; recall that there are only 1’s in
the user-item rating matrix. Several numbers of latent factors
were tested: from 5 to 100, in steps of 5.
ib: an item-based nearest neighbor approach (IB k-NN). This
CF method works in a similar way as the ub approach, but
the similarities are computed between items. In our
experiments we used the Cosine similarity without any
constraint on the neighborhood size; hence, the
neighborhood is limited to the items rated by the user.
cbcf: a hybrid recommendation approach where a user-based
CF strategy is computed using CB user similarities. More
specifically, we compute a Cosine similarity in a similar way
as in the cb algorithm, but between two user profiles instead
of a user and an item profiles. Similarly to the cb algorithm,
cbcf is evaluated by using place, topic and all tags separately.
ipop: a popularity-based recommender. The items
(proposals) with more ratings (comments) are recommended
to the users, without considering any personal information.
4.4 Evaluation methodology and metrics
We focused our evaluation on ranking-based metrics, hence no
rating prediction metrics (such as the Mean Absolute Error and
the Root Mean Square Error) will be reported. The rankings were
generated following the TrainingItems methodology described in
[8], where every item in the training split, except the ones
4
RankSys recommender systems evaluation framework, http://ranksys.org
already seen by the user in training, is considered as a possible
candidate to be part of a user’s final ranking. More specifically,
we followed a 5-fold cross validation strategy to split the dataset
into training and test: 80% of the interactions were randomly
selected to build the training split, and the remainders were used
for the test split.
The reported metrics are the following:
Precision and recall: these metrics measure the amount of
relevant returned items, either normalized by the amount of
items returned (precision) or the amount of relevant items
known for each user (recall).
MAP and nDCG: these two metrics (Mean Average
Precision and normalized Discounted Cumulative Gain)
allow considering differences in the positions of the relevant
returned items.
USC (User Space Coverage): this metric measures the
amount of users who can receive a recommendation (user
coverage). It is important to consider the tradeoff between
USC and recommendation quality (as measured by the
previous metrics).
ISC (Item Space Coverage): this metric measures the
number of different items a recommender is able to
recommend. It is thus related to the diversity of the
recommendations, since the larger its value, the more diverse
the recommendations presented to the users.
For these metrics, we tested several cutoffs, but decided to report
the performance at ranking size of 50 because it was more stable.
4.5 Results
Table 5 shows the results obtained in the evaluation of the
recommendation algorithms using the e-participation dataset
described in Section 4.1.
In general, we observe that using all the tags led to the best
results in cb and cbcf approaches. Although not reported here,
we tested different weighting schemes –binary, TFIDF, and
BM25 [10]–, and for every metric, TFIDF and BM25 achieved
very similar results, and were clearly superior to the binary
weights. Because of this, the values shown in the table
correspond to the TFIDF weighting scheme. Regarding the type
of tags used to create the user and item profiles, we observe that
topic tags outperformed place tags in terms of Precision, MAP,
nDCG and ISC. Place tags, however, provided better coverage
than topic tags, in terms of USC for the cb and cbcf approaches,
and Recall for the cb approach.
The content-based algorithms outperformed the popularity-
based recommender and two of the CF approaches, namely item-
based kNN (ib) and matrix factorization (mf). This could be
attributed to a dataset highly skewed towards content features.
However, the best performing algorithm for most of the metrics
was user-based kNN (ub). This evidences the importance of the
user-item patterns contained in the data, which were not
properly exploited by the other CF algorithms, maybe due to the
very high rating sparsity or the lack of non-unary ratings
between users and items. In contrast, it should be noted that ub
recommendations had a low coverage, for both users (USC) and
items (ISC), and thus despite being precise, they are not diverse
and are generated for a limited number of users. Differently, the
hybrid cbcf approach showed a good tradeoff between
recommendation precision and user coverage, as well as the best
item coverage, i.e., the highest recommendation diversity.
CitRec’17, August 2017, Como, Italy
I. Cantador, A. Bellogín, M. E. Cortés-Cediel, O. Gil
Precision
Recall
MAP
nDCG
USC
ISC
cb
place tags
0.001
0.040
0.005
0.013
0.579
0.087
topic tags
0.001
0.026
0.006
0.010
0.579
0.183
all tags
0.002
0.056
0.013
0.023
0.579
0.241
cbcf5
place tags
0.002
0.011
0.003
0.006
0.577
0.214
topic tags
0.003
0.013
0.005
0.008
0.504
0.230
all tags
0.005
0.019
0.008
0.012
0.563
0.238
cbcf10
place tags
0.002
0.018
0.005
0.008
0.579
0.233
topic tags
0.002
0.021
0.006
0.010
0.513
0.264
all tags
0.003
0.029
0.010
0.015
0.577
0.274
ub5
0.008
0.044
0.017
0.024
0.389
0.243
ub10
0.006
0.059
0.019
0.029
0.467
0.260
ub15
0.006
0.067
0.020
0.031
0.491
0.265
mf5
0.001
0.011
0.001
0.003
1.000
0.044
mf10
0.001
0.016
0.002
0.005
1.000
0.081
mf15
0.001
0.018
0.003
0.006
1.000
0.102
ib
0.002
0.035
0.009
0.015
0.526
0.237
ipop
0.001
0.027
0.006
0.011
1.000
0.004
Table 5: Experimental results.
5 CONCLUSIONS AND FUTURE WORK
Motivated by the need of incorporating personalized information
retrieval and filtering functionalities into e-participation
systems, in this paper we have empirically compared a number
of recommendation approaches for the ‘Decide Madrid’ platform,
where Madrid citizens create, debate and vote proposals for the
city since September 2015.
Assuming that a citizen’s comment on a proposal is a signal of
her interest for the proposal, we have shown that user-based
collaborative filtering heuristics seem to be the best performing
approaches in terms of precision-recall and ranking-based
metrics, such as MAP and nDCG. In contrast, exploiting the tags
assigned by the system users to the proposals, a content-based
approach has achieved the highest coverage values. Finally, a
simple content-based collaborative filtering approach that jointly
uses rating and tag-based information has obtained very good
coverage and diversity values. The real impact of these results
has to be evaluated in a user study.
In this context, the design and evaluation of alternative hybrid
recommendation approaches are left for future work. For
instance, we plan to evaluate the exploitation of other place- and
topic-based user/item profiles inferred from tagging information.
In fact, we have already mapped 100% (11.2%) of the proposals to
their corresponding districts (neighborhoods). Using distance
metrics between districts/neighborhoods may be valuable.
In addition to tags, the citizens’ comments can be further
exploited. Applying NLP and Opinion Mining techniques on the
comments may allow us to determine whether each comment is
in favor or against a particular proposal. With this information,
we would be able to consider binary (like/dislike) ratings, instead
of the unary ratings used in our experiments. Moreover,
evaluating matrix factorization models that exploit content-
based information, conducting experiments on data from the
2016 and 2017 participatory budgeting editions separately, and
considering the ‘whole city’ artificial district differently, are
issues we want to address in the future.
ACKNOWLEDGMENTS This work was supported by the Spanish
Ministry of Economy, Industry and Competitiveness (TIN2016-80630-P).
REFERENCES
[1] Ahlers, D., Driscoll, P., Löfström, E., Krogstie, J., Wyckmans, A. 2016.
Understanding smart cities as social machines. In Proceedings of the 25th
International Conference Companion on World Wide Web, 759–764.
[2] Alawadhi, S., Aldama-Nalda, A., Chourabi, H., ... & Walker, S. 2012. Building
understanding of smart city initiatives. In Proceedings of the 11th
International Conference on Electronic Government, pp. 40–53.
[3] Álvarez-Sabucedo, L., Soto-Barreiros, R., Santos-Gago, J. M., Fernández-
Iglesias, M. 2012. A hybrid semantic driven recommender for services in the
eGovernment domain. In Proc. of the 2nd Intl. Conference on Digital
Information and Communication Technology and its Applications, pp. 409–414.
[4] Andrés, J., Bonivento, H. 2009. Problemas institucionales de la
participación ciudadana : Análisis conceptual y aplicación al caso
colombiano. XXII Concurso del CLAD sobre la reforma del Estado y
modernización de la Administración Pública, pp. 1–23.
[5] Ayachi, R., Boukhris, I., Mellouli, S., Ben Amor, N., Elouedi, Z. 2016.
Proactive and reactive e-government services recommendation. Universal
Access in the Information Society 15(4), 681–697.
[6] Baldassarre, C., Cremaschi, M., Palmonari, M. 2013. Bridging the gap
between citizens and local administrations with knowledge-based service
bundle recommendations. In Proceedings of the 24th International
Conference on Database and Expert Systems Applications, pp. 157–161.
[7] Basañes, F., Bonnefoy, J. C., Yitani, J. E., … Schmuck, E. 2015. Liderando el
desarrollo sostenible de las ciudades. Banco Iberoamericano de Desarrollo.
[8] Bellogín, A., Castells, P., Cantador, I. (2011). Precision-oriented evaluation
of recommender systems: An algorithmic comparison. In Proceedings of
the 5th ACM Conference on Recommender Systems, pp. 333–336.
[9] Buccoliero, L., Bellio, E. 2016. Marketing and citizen web empowerment: An
index for benchmarking Italian municipalities’ web strategies. In Proc. of the
9th Int. Conference on Theory and Practice of Electronic Governance, pp. 17–26.
[10] Cantador, I., Bellogín, A., Vallet, D. 2010. Content-based recommendation in
social tagging systems. In Proceedings of the 4th ACM Conference on
Recommender Systems, 237–240.
[11] Coleman, S. 2008. Doing IT for themselves: Management versus
autonomy in youth e-citizenship. In W. Lance Bennett (Ed.), Civic life
online: Learning how digital media can engage youth, pp. 189–206.
[12] Hernández-Bonivento, J. A. 2011. La gobernanza local y el fortalecimiento
institucional. GIGAPP-IUIOG Estudios Working Papers, num. 2011-08.
[13] Hu, Y., Koren, Y., Volinsky, C. 2008. Collaborative filtering for implicit
feedback datasets. In Proceedings of the 2008 8th IEEE International
Conference on Data Mining, pp. 263–272.
[14] Kavanaugh, A., Krishnan, S., Pérez-Quiñones, M., Tedesco, J., Madondo,
K., Ahuja, A. 2014. Encouraging civic participation through local news
aggregation. Information Polity 19(1, 2), 35–56.
[15] Macintosh, A. (2004). Characterizing e-participation in policy-making. In
Proc. of the 37th Hawaii Intl. Conference on System Sciences, pp. 10–pp.
[16] Marsal-Llacuna, M. L., De la Rosa-Esteva, J. L. 2013. The representation
for all model: An agent-based collaborative method for more meaningful
citizen participation in urban planning. In Proc. of the 13th International
Conference on Computational Science and its Applications, pp. 324–339.
[17] Meijer, A., & Bolivar, M. P. R. 2015. Governing the smart city: A review
of the literature on smart urban governance. International Review of
Administrative Sciences 82(2), 392–408.
[18] Nam, T., Pardo, T. A. 2011. Conceptualizing smart city with dimensions
of technology, people, and institutions. In Proceedings of the 12th Annual
International Digital Government Research Conference, pp. 282–291.
[19] Nelimarkka, M., Nonnecke, B., Krishnan, S., ... Patel, J. 2014. Comparing
three online civic engagement platforms using the “spectrum of public
participation” framework. In Proceedings of the 3rd Internet, Policy, and
Politics Conference.
[20] Organisation for Economic Co-operation and Development (OECD).
2001. Citizens as partners: Information, consultation and public
participation in policy-making. OECD Publishing.
[21] Organisation for Economic Co-operation and Development (OECD).
2003. Promises and problems of e-democracy; Challenges of citizen online
engagement. OECD Publishing.
[22] Terán, L. F., Meier, A. 2010. A fuzzy recommender system for eElections.
In Proceedings of the 1st International Conference on Electronic Government
and the Information Systems Perspective, pp. 62–76.
[23] Salim, F., & Haque, U. 2015. Urban computing in the wild: A survey on
large scale participation and citizen engagement with ubiquitous
computing, cyber physical systems, and Internet of Things. International
Journal of Human-Computer Studies 81, 31–48.