A Fuzzy-Based Platform for Stimulating Citizens’Participation

Luis TeranInformation Systems Research Group

University of FribourgFribourg, Boulevard de Perolles 90

Email: luis.teran@unifr.ch

Abstract—Voting advice applications, social networks and vir-tual communities have become a hot topic in today’s society. Suchtechnologies could also improve democratic processes, increasecitizens’ interest in political issues, enhance participation, andrenew civic engagement. However, the difficulty of finding othercitizens or parties that share common goals is still a barrier.In this paper, the SmpartParticipation platform is introduced. Ituses a fuzzy-based recommender system architecture for recom-mending political parties, candidates, and the creation of virtualcommunities. The platform, uses the so-called fuzzy profiles forboth: citizens and candidates to define political orientation andinterest on different issues. Afterwards, the recommendationengine computes fuzzy clusters for the recommendation. Finally,citizens can evaluate the top-n politicians from different politicalparties in an election process and/or build up communitiesthat share their common interests. The recommendation engineprovides to users, a bi-dimensional political/issue-based landscapeto better understand their proximity to politicians or issues.

Keywords-voting advice applications; recommender system;community building; eParticipation; eDemocracy; eCommunity;fuzzy clustering;

I. INTRODUCTION

In recent years, the use of voting advice applications (VVAs)have become very popular. Thus, the advice given is of greatpolitical importance for opinion formation, decision-makingand voting behaviour.

In the work of Wagner and Ruusuvirta [1], the recommen-dations given by twelve VAAs in seven European countries isexamined. It describes problems of effectiveness at establish-ing party positions, which can lead to faulty recommendations.

In this paper, an new approach for establishing party po-sitions is presented, which combines Web-based VAAs withfuzzy logic. With fuzzy classification, a politician or citizencan belong to more than one class with differing degrees ofmembership. This notion of membership not only provides abetter description of political parties but also helps users makedifferentiated decisions in eElection processes (Teran andMeier [2], Teran et al. [3]) and when building communities.

The fuzzy-based recommendation engine used in the Smart-Participation project for citizens and politicians uses datafrom the smartvote project1, which corresponds to the Swissnational elections in 2007. The smartvote web site is a political

1smartvote project: http://smartvote.ch

issue-matching VAA with analytical tools for evaluating andanalyzing political positions (Fivaz and Schwarz [4]).

In the fuzzy-based prototype, a modified fuzzy c-meansalgorithm and the Sammon mapping technique are used forclustering political parties, citizens and political issues. Ap-plicants can therefore use the fuzzy recommender systems toevaluate the top-n neighbors (politicians and/or citizens) with afuzzy similarity range, clusters of political parties and citizens,or build up new political communities in a political/issue-basedlandscape.

This research paper is structured as follows: Section IIdescribes the SmartParticipation project. Section III, delineatesthe fuzzy-based recommender system architecture. It consistsof an overview and the description of building blocks, suchas user profile generation, the recommender engine, fuzzyclustering, the top-n recommendations, and political commu-nity building. Section IV proffers concluding remarks andsuggestions for future research.

II. THE SMARTPARTICIPATION PROJECT

The SmartParticipation project intends to provide citizenswith a simple and innovative alternative based on fuzzyclusters for monitoring and evaluating the performance andobjectives of political actors. Additionally, this applicationallows citizens to create virtual communities based on theirprofiles, such as new political parties, thematic groups andcivic networks, participating in national issues by openingchannels of discussion and debate through the use of infor-mation and communication technologies (ICTs) and Web 2.0.

Two main tools have been designed in the SmartPartici-pation application: fuzzy clustering and virtual communities.Fig. 1 illustrates the use of such tools on different topics relatedto eCollaboration, eDemocracy and eCommunity.

The current recommendation engine, described in this paper,has been designed for eDemocracy and eCommunity pro-cesses. Further work will also cover eCollaboration.

III. FUZZY-BASED RECOMMENDATION ENGINE

Although collaborative filtering-based approaches are morewidely used for recommender systems in eCommerce to sug-gest items that a customer is presumably going to buy, theyare only suitable in the repeat-appeared scenario, which isdescribed by Vozalis and Margaritis [5].

2353978-1-4577-1979-0/12/$26.00 ©2011 IEEE

SmartParticipation

eDemocracyeCollaboration eCommunity

→ Current work → Future work

Virtual OrganizationsForms of CooperationCollaborative Working

Thematic TopicsDiscussion ForumsDecision Making AidsPublic MemoryPolitical EvaluationPolitical Control

Civic NetworksMatchmaking SystemsRecommender SystemsNew Political PartiesThematic Groups

Figure 1: The SmartParticipation project.

Recommender systems for eDemocracy and the creation ofcommunities must also be suitable in the one-and-only itemsscenario, in which the recommendation target is a uniqueitem/event (e.g., a voter v wants to receive a recommendationof n candidates that are close to his preferences in an electionE).

The SmartParticipation project uses a fuzzy recommendersystem prototype (FRSP) developed to display the resultsof a recommendation. Three output options were developed:fuzzy cluster analysis, top-N recommendations and communitybuilding tools, as shown in Fig. 2.

A. Architecture Overview

The recommendation procedure is divided into three steps:In the first step, the voters (users) and candidates must createtheir profiles using a fuzzy interface, which is a convenienttool used to determine the level of agreement, disagreement,and relevance for each specific question. The fuzzy profilesare stored in a database.

In the second step, once all necessary profiles have beencreated, the user selects the recommendation target and thetype of output (top-N recommendation, fuzzy cluster analysisor political community).

In the final step, once the recommendation engine hascomputed all the information, the user receives the recom-mendation in the pre-established format.

The architecture of the fuzzy recommendation approach ispresented in Fig. 2. Each element is presented in more detailin the following sections.

B. User Profile Generation

In order to provide a recommendation, voters (users) andcandidates must generate a profile that describes their pref-erences using a fuzzy interface to complete a questionnaireregarding political issues (each question has different possibleresponses).

The fuzzy interface is a convenient tool used to determinethe level of agreement/disagreement and relevance for a spe-cific question. Unlike other similar tools, it provides a highernumber of possibilities for each citizen/candidate to answerthe questions. The interface is designed to be as intuitive andconvenient as possible for users.

In addition to the fuzzy interface, the system contains aprofile representation, so-called fuzzy profile (FP), which is a

Fuzzy Interface

Voter (User) / Candidate

Database

❶ Fuzzy Profile Generation

RecommendationEngine

❷ Ask for Recommendation

❸ Receive

Recommendation

Fuzzy ClusterAnalysis

Fuzzy Clustering Political orIssue-based Community

Top-N Recommendation

Figure 2: Fuzzy-based recommender system architecture

fpc1,11

fpc1,22

fpc1,32

FP1 = (fpc1,11, fpc1,2

2 , fpc1,32)

I1(a1=1) I2(a2=2)

Figure 3: Fuzzy profile (FP)

multi-dimensional Euclidean space with a total of n elements,which is defined as follows:

FPi = (fpc1i,1, fpc1i,2, ..., fpc

1i,a1︸ ︷︷ ︸

I1(a1)

, fpc2i,a1+1, ..., fpc2i,a1+a2︸ ︷︷ ︸

I2(a2)

, ...

..., fpcki,a1+a2+...+ak−1+1, ..., fpcki,a1+a2+...+ak−1+ak︸ ︷︷ ︸

Ik(ak)

)

where FPi is the FP of user i, fpckij is the j-th fuzzy profilecomponent (fpc), and Il represents a specific issue (with a totalnumber of issues = k) . Each issue Il has a set of positivenumber of questions defined by al, and the total of questionsof all issues is equal to n, such that:

k∑l=1

al = n

Fig. 4 show an instance of FP, which has three fpc andtwo issues (I1 and I2). Additionally, I1 has only one question(a1 = 1), while I2 has two questions (a2 = 2).

The fpc hast two components: agreement (agrkij → [−1, 1])and relevance (relkij → [0, 1]). Each fpc corresponds to theanswer of a question that belongs to a specific issue Il, it isdefined as follows:

2354

AgreeDisagree

-1 10

Relevant

0.5

1

Fully disagree = -1* || (-1, 1) || = -1.41

Fully agree = 1* || (1, 1) || = 1.41

fpci,jk = -1* || (-1, 0.5) ||

= -1.11

Indifferentand Irrelevant

Figure 4: Fuzzy profile component (fpc)

fpckij = α(agrkij) ∗ ||(agrkij , relkij)||

and

α(agrkij) =

{1 if agrkij ∈ [0, 1]−1 if agrkij ∈ [−1, 0)

where agrkij is the j-th component of agreement on the k-thissue, and relkij is the j-th component of relevance on the k-thissue. Fig. 4 show an instance of fpckij .

Additionally, Fig. 4 shows the references that are usedfor specifying: full agreement, full disagreement and indif-ference/irrelevance. They are used for building communitiesbased on issues, and they discussed in more detail in sec-tion III-F.

C. Recommendation Engine

The recommendation engine is based on the generation offuzzy clusters using a modified fuzzy c-means algorithm.

Once the profiles are generated, the next step is to ask fora recommendation. At this point, the user selects a particularevent and the type of recommendation (top-N recommenda-tion, fuzzy clustering analysis or political/issue-based commu-nity). The request is sent to the recommendation engine, whichprocesses the query.

To provide a graphical representation of results that userscan easily analyze, the recommendation engine transforms thehigh-dimensional space of profiles to a lower dimensionalspace (bi-dimensional), which reduces the complexity of dataanalysis. The recommender engine uses a mapping methodoriginally proposed by Sammon [6] that attempts to preserveinter-pattern distances.

D. Fuzzy Cluster Algorithm

Once the profiles are mapped to a low-dimensional spacewith the Sammon mapping technique, the recommendationengine generates fuzzy clusters by using a modified FCMalgorithm, which requires two main inputs: the number ofclusters, and a matrix of cluster centers. For this reason, priorknowledge of the dataset is required.

The modified fuzzy c-means algorithm is a two-step iterativeprocess that is defined as follows: First, set the input variablesc (number of clusters is equal to number of political parties

or number of issues), m (level of fuzziness), ε (terminationcriterion, normally ε ∈ [0, 1]), and t (type of clustering, it couldbe either “Political Party” or “Issue-based” ).

First Case: If type is equal to “Political Parties”, therecommendation engine considers the number of clusters to beequal to the number of political parties. In the second step thealgorithm sets an iteration number k = 0. After that, generatea matrix of cluster centers ~Y (k) taking the mean average ofanswers from all candidate in the same political party (Pi).Then, given the initial matrix ~Y (k), compute the fuzzy partitionmatrix ~U (k).

Finally, using a repeat-until loop, update ~Y (k+1) using ~U (k)

and then update ~U (k+1) using ~Y (k+1). Repeat this processuntil the termination criterion is reached (|~U (k+1)−~U (k)| ≤ ε).

The termination criterion could also be: a predefined numberof iterations, or a condition that updates the centers, only ifthe number of candidates in a political party are majority inthe cluster, otherwise, the center does not update.

Second Case: If type is equal to “Issue-based”, thealgorithm sets the center of all issues for both: full agreementand full disagreement (refer to Fig. 4). After that, and given theinitial matrix ~Y (k), compute the fuzzy partition matrix ~U (k).The reason of taking only one iteration is that the centers arenot expected to move.

The modified fuzzy c-means is presented in Algorithm 1.

Algorithm 1 FCM Modified

Input: c,m, ε, tOutput: ~U (k+1), ~Y (k+1)

1: if t = Political − Party then2: Set iteration number: k ← 03: for i = 1 to c do4: yi ← mean average of answers from the i-th Political

Party (Pi)5: end for6: Compute ~U (k) ← ~Y (k)

7: repeat8: Update ~Y (k+1) ← ~U (k)

9: Update ~U (k+1) ← ~Y (k+1)

10: until |~U (k+1) − ~U (k)| ≤ ε11: else if t = Issue−Based then12: for i = 1 to c do13: yi ← full agreement on issue Ii14: yi+1 ← full disagreement on issue Ii15: end for16: Compute ~U (k) ← ~Y (k)

17: end if18: return ~U (k+1), ~Y (k+1)

The outputs of the modified fuzzy c-means algorithm are:a fuzzy partition matrix ~U (k+1) that contains the membershipdegree of voters (users) and candidates/citizen with respect toeach cluster, and a matrix of cluster centers ~Y (k+1).

The prototype developed in this paper displays, in a bi-dimensional map, the locations of the voter (user) and the

2355

candidates (labeled by political parties), the clusters that aregenerated according to each political party, and the percentageof the closeness of the voter (user) to each cluster. Fig. 5 showssome examples of the prototypes interface, for both references:political parties and issue-based.

Fig. 5a displays the formation of clusters with a clearconcentration of candidates from the same political party. Itshows that the closest political party with respect to the voter(user) is the Federal Democratic Union (66%), followed bythe Social Democratic Party (22%) and finally the RadicalDemocratic Party (12%).

E. Top-N Recommendations

The top-N candidates/citizens similar to voter (user) v aregenerated by using the bi-dimensional profiles. The distancesof all candidates, with respect to voter (user) v, are computed,and the N closest candidates are displayed. The similarity per-centage (Svci(%)) of a voter (user) v and the i-th candidate (ci)is computed using the most distant candidate or citizen (dmax)as a reference. The computation of similarity percentage is:

Svci(%) = 100− (100 ∗ dvcidmax

)

where dvci is the distance between voter (user) v and the i-thcandidate/citizen.

The prototype developed in this paper displays the locationof a voter (user) and candidates (labeled by political parties),the clusters generated according to each political party (witha percentage of closeness of the voter (user) to each cluster),and the N closest candidates labeled with the percentage ofproximity to the voter (user).

Fig. 5b shows the results with the same dataset used inFig. 5a. It shows the formation of clusters by political partyand the top-10 candidates close to the voter (user), togetherwith the similarity percentages.

F. Building Communities

The recommendation engine presented in this paper couldalso be used during eDiscussion and ePosting, which ispresented in the work of Meier [7], with the creation ofvirtual communities, allowing citizens to interact throughspecific media, potentially crossing geographical and politicalboundaries in order to pursue mutual interests or goals. Theuse of the user-friendly bi-dimensional interfaces, could helpvoters (users) to establish which citizens are the most similaraccording to their preferences and tendencies (profiles).

To create political communities, the recommendation engineuses the datasets of citizens together with the dataset ofcandidates. The prototype developed in this paper has twotypes of references for centes of cluster: political parties andissue-based.

First Case: In the case of using political parties asreferences, the recommendation engine transforms the high-dimensional profiles into a bi-dimensional space. Secondly, inorder to compute the fuzzy clusters, only the bi-dimensional

profiles of candidates and the voter (user) looking for the rec-ommendation are used. Once the fuzzy clusters are computed,the datasets are merged (voter, citizens and candidates) anddisplayed in a bi-dimensional map.

In Fig. 5c, not only the candidates but also the citizensinvolved in the system are included in the bi-dimensional map.The citizens are represented by black squares, and for thisexperiment, the 20 closest citizens are represented by filledblack squares.

Second Case: In the case of using issues as references,the recommendation engine transforms the high-dimensionalprofiles into a bi-dimensional space. Secondly, in order tocompute the fuzzy clusters, only the bi-dimensional referencesof issues are taken for both: full agreement and full disagree-ment. Once the fuzzy clusters are computed, the datasets aredisplayed (voter, and citizens) in a bi-dimensional map.

Fig. 5d shows how close is the user with respect to issues3 and 5 (the issues are taken from the smartvote dataset). Itshows that the user is 45% to the full agreement on issue3 and 38% to full disagreement of issue 5. The citizens arerepresented by black squares, and for this experiment, the 20closest citizens are represented by filled black squares.

IV. CONCLUSIONS AND OUTLOOK

In this research, a recommender systems architecture foreGovernment, used in the SmartParticipation project, has beenproposed. The Web-based recommendation engine can beused to visualize differentiated clusters of politicians as wellas of citizens. It therefore supports collaboration, eElectionprocesses for candidates, building processes for political com-munities that share common objectives, and civic participation.

The recommender system proposed can be used for eCol-laboration, eDemocracy, and eCommunity. Based on a fuzzyclustering approach, it computes similarities between citizensand politicians in a multi-dimensional space. The Sammonmapping technique allows for a better understanding and eval-uation of the relationships among citizens and/or politiciansusing a bi-dimensional graphical interface.

The creation of political communities and social networksamong citizens allows for interaction and participation throughsocial media, potentially crossing geographical and politicalboundaries. Contacting people with similar political profiles,building exchange platforms, and stimulating participation willenrich the information and knowledge-based society in thefuture.

The recommender system approach presented in this paperdiffers from collaborative filtering methods in that they arebased on past experiences. It is also suitable in the one-and-only scenario, in which events such as election processes occuronly once, and their participants (candidates and/or citizens)cannot be considered unique, since their presence at suchevents and their way of thinking can vary over time.

In future work, the SmartParticipation project could beused to evaluate whether candidates really act the way theyclaim they will. The FRS could display their location inthe bi-dimensional map as candidates and show their moving

2356

Voter (User)Center SDP 22%Candidates SDP

Candidates RDP

Center RDP 12%

Center FDU 66%Candidates FDU

(a) Fuzzy cluster analysis graphical interface

Voter (User)Center SDP 22%Candidates SDP

Candidates RDP

Center RDP 12%

Center FDU 66%Candidates FDU

70%

91%

86%

85% 86%

80%

66%

66%

70%

71%

Bold figures: Top-10 candidates

(b) Top-N recommendation graphical interface

Voter (User)Center SDP 22%Candidates SDP

Candidates RDP

Center RDP 12%

Center FDU 66%Candidates FDUTop-20 Citizens

Citizens

(c) Political communities graphical interface(reference: Political Parties)

Voter (user)Issue 5 (Fully Agree) 8%Issue 5 (Fully Disagree) 38%Issue 3 (Fully Agree) 45%Issue 3 (Fully Disagree) 9%20 Closest Citizens

Citizens

(d) Political communities graphical interface(reference: Political Issues)

Figure 5: Recommendation output.

positions during their political engagement as elected officials,allowing voters to easily understand politicians’ behavior.

The recommender system approach presented in this paper,the fuzzy interface, and the graphical interfaces must beevaluated by citizens. In the case of the algorithms used, acomparison with different methods for dimensionality reduc-tion and clustering algorithms will be performed.

ACKNOWLEDGMENT

The author would like to thank the members of the smart-vote project (smartvote.ch), the members of the Fuzzy Mar-keting Methods Research Center (www.FMsquare.org), andthe Information System Research Group at the Universityof Fribourg (diuf.unifr.ch/main/is) for contributing valuablethoughts and comments. Special thanks to Andreas Ladner,and Jan Fivaz for their support and help using the smartvotedatabase to test the prototype.

REFERENCES

[1] M. Wagner and O. Ruusuvirta, “Faulty recommendations? party positionsin online voting advice applications,” SSRN eLibrary, 2009.

[2] L. Teran and A. Meier, “A fuzzy recommender system for e-elections,”in Electronic Government and the Information Systems Perspective, ser.LNCS, vol. 6267. Bilbao: Springer, September 2010, pp. 67–76.

[3] L. Teran, A. Ladner, J. Fivaz, and S. Gerber, Fuzzy Methods for CustomerRelationship Management and Marketing: Applications and Classifica-tions. IGI Global, January 2012, ch. 6, pp. 115–138.

[4] J. Fivaz and D. Schwarz, “Nailing the pudding to the wall: E-democracyas catalyst for transparency and accountability,” in International Confer-ence on Direct Democracy in Latin America, Buenos Aires, Argentina,March 2007.

[5] E. Vozalis and K. G. Margaritis, “Analysis of recommender systems’algorithms,” in Proceedings of the Sixth Hellenic-European Conferenceon Computer Mathematics and its Applications - HERCMA 2003, Athens,Greece, September 2003.

[6] J. W. Sammon, “A nonlinear mapping for data structure analysis,” IEEETransactions on Computers, vol. c-18, no. 5, pp. 401–409, May 1969.

[7] A. Meier, eDemocracy & eGovernment - Maturity Levels of a DemocraticKnowledge Society. Berlin: Springer-Verlag, 2009.

2357