=Paper=
{{Paper
|id=Vol-2578/BigVis6
|storemode=property
|title=The new Spanish political scenario: Twitter graph and opinion analysis with an interactive visualisation
|pdfUrl=https://ceur-ws.org/Vol-2578/BigVis6.pdf
|volume=Vol-2578
|authors=Pelayo Quirós Blanca,Rosario Campomanes-Álvarez
|dblpUrl=https://dblp.org/rec/conf/edbt/BlancaC20
}}
==The new Spanish political scenario: Twitter graph and opinion analysis with an interactive visualisation==
https://ceur-ws.org/Vol-2578/BigVis6.pdf
The new Spanish political scenario: Twitter graph and opinion
analysis with an interactive visualisation
Pelayo Quirós∗
Blanca Rosario Campomanes-Álvarez∗
pelayo.quiros@ctic.es
charo.campomanes@ctic.es
CTIC Technological Centre
Gijón, Asturias, Spain
ABSTRACT Furthermore, these applications are usually centred in a partic-
Twitter, as a social network, has been used extensively as a way ular field, given the raw daily volume, as it has been previously
to express users opinion about a wide range of topics. Particu- mentioned. Among others, economics ([33], [25], [10]), television
larly, politics have been one of the main issues in the last years, and cinema ([2], [12], [30]), or sports ([28], [36], [31]) have been
and consequently, its study in Twitter can lead to valuable con- widely studied. However, this paper is focused on the application
clusions. of Twitter analytic to politics, which has already been tackled
The methodology presented in this paper is based on the ex- as well ([15], [16], [4], [5]). Particularly, this analysis is devoted
traction of the followers network of the main politicians in Spain to the Spanish political situation in Twitter prior to the second
prior to the general election of November of 2019, providing a general election in 2019, which took place the 10th of November.
suitable graph analysis and visualisation. Hence, closeness be- The main motivation of this approach is based on the re-
tween different political parties can be spotted with respect to cent evolution of the Spanish political scenario (http://www.
their common followers. This is motivated by the new politi- infoelectoral.mir.es/). It has been a two-party system since the
cal scenario, given that in the last five years, it derived from general election of 1989, where PSOE and PP received most of
a bipartisan system to a plural one, with six relevant national the votes. Particularly, the total percentage of votes obtained by
parties. these two parties ranged from 65% in 1989, to 73% in 2011, getting
Additionally, information about each political party is pre- their highest combined percentage in 2008 with a total over the
sented interactively with respect to the political profile of those 83% of votes. However, in 2015 two new parties irrupted, Unidas
who write about them. Reciprocally, this tool also provides infor- Podemos and Ciudadanos, reducing dramatically the bipartisan
mation about the target of this party identified profiles. percentage to 50% of votes. Another two new political parties
were added to this scenario in 2019, VOX and Más País, turning
KEYWORDS the sum of votes of PSOE and PP, for the first time since their
coexistence, into values under the 50% (45% in April, 48% in No-
Twitter, political analysis, social networks, graphs, data visuali-
vember). Additionally, regionalist parties have strengthen their
sation
presence in their respective areas, with a sum of 10% and 11% of
the total votes in both general elections in 2019, respectively.
1 INTRODUCTION As a consequence, with such a variety of political parties, it is
hard to assess the closeness between certain parties. Even though
Twitter is a microblogging social network based on the publi-
the leaders may point out their preferences, their voters may have
cation of short messages of 280 characters at most. It has been
a different opinion, and as such, analysing their closeness with
used extensively as an opinion network with daily volumes of
respect to their supporters is a must in order to understand the
generated messages of about 500 million tweets, with 139 million
underlying structure of the actual political Spanish scenario.
daily active users (https://business.twitter.com/). Every user has
To analyse such closeness with respect to supporters, all the
the option to follow other accounts, so the content published
followers of the most remarkable politicians associated to the ten
by them is shown to their followers. The different actions with
most relevant political parties have been downloaded via Twitter
respect to each Twitter message or tweet are sharing it with your
API, including the two bipartisan parties (PSOE, PP), the national
followers (retweet), marking it as favourite or answering the
ones of recent creation (Unidas Podemos, Ciudadanos, VOX, Más
tweet. Furthermore, every individual can mention another user
País), and the four most relevant regionalist parties (ERC, Junts
by using the account name preceded by an "@".
Per Catalunya, PNV, EH Bildu).
Consequently, many approaches have been studied concerning
All this information has been processed in order to put it
different types of insights. Sentiment analysis has been one of
together as a graph or network. The main goal of this work is
the most relevant ([29], [22], [21], [1], [14]), as well as graph
to better understand the relationship between user and political
analytic of the connection network among Twitter users ([11],
Twitter accounts, as well as discovering insights about the new
[27], [26]) or of the graph generated by their communication
and uncertain political Spanish scenario. This could be done by
through mentions, answers and retweets ([34], [8], [37]).
means of Social Network Analysis (SNA) techniques due to the
∗ Both authors contributed equally to this research. nature of the used data.
SNA is used for measuring and analyzing the structural prop-
© 2020 Copyright for this paper by its author(s). Published in the Workshop Proceed- erties of networks of interdependent dyadic relationships. One of
ings of the EDBT/ICDT 2020 Joint Conference (March 30-April 2, 2020, Copenhagen, the core assumptions of SNA is that the patterns of these relation-
Denmark) on CEUR-WS.org. Use permitted under Creative Commons License At-
tribution 4.0 International (CC BY 4.0) ships can have important effects on individual and organizational
behaviour, constraining or enabling access to resources, and ex-
posure to information and behaviour [17]. Apart from this, one
of the key elements that characterizes modern SNA is the use of
visualisations of complex networks. Innovators in information
visualisation have also contributed to helping users to discover
patterns, trends, clusters, gaps, and outliers, even in complex
social networks [32].
In order to complement this study, an analysis of the different
messages directed to each of the candidates in Twitter has been
developed, obtaining the political typology of those who speak
about them.
This paper is structured as follows. Section 2 presents the
whole proposed methodology. In Section 3 the development of
the application, along with visual examples, is provided. Section
4 is devoted to conclusions and future work.
Figure 1: Diagram of the proposed methodology
2 PROPOSED METHODOLOGY
As previously said, the recently appeared new parties in Spain from and to the selected party, with the previously obtained
have added important and unknown changes to the political profiling set. Both graphical outputs are presented in a website
tendencies in the public opinion. This constitutes a new scenario that allows a joint visualisation. This web application is based on
in which voters have changed their preferences from only two the Node.js [13] engine. It has been developed in order to carry
options to a system with multiple parties, with a significant rise out a post visual data exploration and an interactive analysis of
of nationalist ones. Understanding and visualising these new the network.
political trends and how the vote-flow can vary from one party
to another or how parties are related between them, arises as 2.1 Data Extraction and Processing
an interesting task for interpreting the new complex political Data extraction is carried out by using the official Twitter API
scenario in Spain. (https://developer.twitter.com/), which allows to obtain a wide
The tool presented here is based on the study of the large range of information from Twitter, such as timelines, tweets
network that represents the different connections among Twitter mentioning a certain term, hashtag or account, or the users that
users, where a relationship between two given individuals is follow a given individual. In this paper, the proposed approach is
not necessarily reciprocal, which leads to a directed graph study based on the latter one. Consequently, it is mandatory to decide
corresponding to what users follow each individual. which accounts should be analysed for every political party. For
Particularly, this work has been devoted to analysing the Span- each one, the official account from the party, their corresponding
ish political Twitter environment, previous to the general election leaders, as well as additional remarkable politicians have been
that took place the 10th of November of 2019. The ten main politi- monitored. In Table 1 a list with each political party and the
cal parties in Spain have been monitored: PSOE, PP, VOX, Unidas number of accounts analysed for each one is presented.
Podemos, Ciudadanos, Más País, ERC, Junts Per Catalunya, PNV
and EH Bildu. The first six are national parties, while the remain- Table 1: Summary of accounts analysed for each political
ing ones are only present in certain areas (Catalonia for ERC and party
Junts Per Catalunya, Basque Country for PNV and EH Bildu).
Hence, the proposed approach aims to perform an interactive Political party Number of accounts
network by: 1) extracting and processing data based on political PSOE 11
interactions from Twitter, 2) from this data, creating a graph and PP 11
extracting communities from it, 3) applying a layout algorithm VOX 10
to align nodes according to their community membership; and 4) Unidas Podemos 11
producing a two-dimensional graph visualisation in a web-based Ciudadanos 11
platform that supports pan, zoom navigation and an additional Más País 11
data analysis of the relevant nodes by means of the interaction ERC 9
with them. Junts Per Catalunya 8
The whole architecture of the proposed system is depicted in PNV 7
Figure 1. This diagram shows two connected lines of action. The EH Bildu 7
first one is based on the extraction of followers of each politician TOTAL 96
from the ten political parties, which are afterwards clustered and
filtered, obtaining a set of followers reduced and processed, in As it can be observed, the number of accounts differs slightly
order to apply the different graph methodologies with the open- among them. This is due to the lack of relevant and visible ad-
source SNA software Gephi (version 9.1) [6]. Furthermore, the ditional politicians for the minor parties. Globally, 96 political
original followers dataset is used to obtain political profiles with accounts have been selected.
respect to that ten parties. On the other hand, a network consists of two components,
Secondly, a daily download of tweets that mention the six a list of the actors which compose the network, and a list of
national political parties and their respective leaders is conducted. relations, i.e., the interactions between actors. As part of a math-
These messages are processed to obtain the profile of opinion ematical object, actors will then be called vertices or nodes, and
relations will be denoted as edges. Based on graph theory, a net- aggregation of the network based on the refined partition,
work corresponds to a graph, which is built by vertices or nodes using the non-refined partition to create an initial partition
as well as edges connecting vertices between them. This way of for the aggregate network.
representing data is appropriate for scenarios like the proposed
one, which involves a high number of connections and can not Once the community detection is performed, the next step
be accurately understood and represented by using traditional consists of evaluating the results obtained by the community-
graphics due to its complex structure. detection algorithms. The selected metric for determining the
Taking into account the extracted data and the graph theory, quality of a community is the modularity [24]. This measure is
a directed network can be built from the generated dataset. Two based on the idea that a random graph is not expected to have a
types of nodes are defined. The first type of them refers to the cluster structure. Therefore, the possible existence of clusters is
political accounts, whereas the other one determines the user revealed by the comparison between the actual density of edges
accounts, i.e., individuals following political accounts. Regarding in a subgraph and the density that would be expected in a random
the interaction between nodes, an edge represents the relation- subgraph.
ship that is created when a user follows a political account. Furthermore, other centrality measures can be calculated in
Once the initial information is downloaded, as it has been order to provide the importance, or influence, in a social network.
previously explained, it is necessary to store and process it to For instance, the Degree Centrality indicates what accounts are
apply the different graph processing, visualisation and analysis the most followed, the Eigenvector Centrality shows which are
techniques. the most influential accounts, and the Betweenness Centrality
Given the complexity granted by data dimension, it is manda- detects who/which are the users/accounts controlling the infor-
tory to simplify it, so it can be manageable by usual graph tools. mation flow [38].
In order to do so, the original data is modified by aggregating
every user who follows exactly the same political accounts into a
common cluster, whose node size is associated to the number of 2.3 Graph visualisation
individual that it represents. Furthermore, each node has an edge Once the communities are calculated, a colour for each of them
to each political account that this cluster of accounts follows. (nodes and edges in a particular community) are assigned to
Simultaneously, it is developed a methodology that determines better distinguish the structure of the different groups inside the
which users are really interested in one or many political parties, graph. In addition, the node size is also determined by firstly
while applying a filter that removes those whose interest is weak, distinguishing between political and user nodes. The size of the
circumstantial or too spread among several political parties. This political nodes is modified based on the number of followers they
information is complementary used to the graph-associated one, have, i.e., political nodes with a higher in-degree present a larger
in order to enrich the provided visualisation. size than political nodes with a lower in-degree. On the other
hand, the size of the user nodes will be larger when these nodes
2.2 Graph Processing: Community Detection group together more individuals.
Algorithms Regarding the information of each node, the visualisation of
Community detection algorithms in graphs are used for identify- the graph also involves showing the data related to each political
ing groups of similar individuals in order to understand user in- node such as the name of the political party or the photograph
teractions and behaviours. In addition, some of these behaviours of the politician. This could be useful, for instance, to rapidly
are only observable into a group and not on an individual level. distinguish if nodes of different political parties belongs to the
This is because individual behaviours could easily change, but same community or in which position of the political scenario
collective behaviours are more robust to changes [7]. These algo- (the calculated graph) they are located.
rithms are also useful to define the communities of highly related After establishing the appearance of the whole network, a
nodes as well as visualising their relations to other communities. layout algorithm should be applied for drawing the graph in
In this work, the algorithms considered for identifying differ- an aesthetically way as well as differentiating the communities
ent communities in the graph, are the following: within the graph.
Their purpose is to position the nodes of a graph in a two-di-
• The Louvain method [9]: This algorithm detects communi-
mensional space so that all the edges are of more or less equal
ties in networks by maximizing a modularity score for each
length and there are as few crossing edges as possible. This is
community, where the modularity quantifies the quality
done by assigning forces among the set of edges and the set of
of an assignment of nodes to communities by evaluating
nodes, based on their relative positions, and then using these
how much more densely connected the nodes within a
forces either to simulate the motion of the edges and nodes or to
community are, compared to how connected they would
minimize their energy [3].
be in a random network. This method is one of the fastest
The considered layout-graph algorithms for modelling the
modularity-based algorithms, and works well with large
shape of the graph are the following:
graphs. It also reveals a hierarchy of communities at dif-
ferent scales, which can be useful for understanding the
• ForceAtlas2 [18]: This algorithm handles large networks
global functioning of a network.
while keeping a very good quality. Nodes repulsion is
• The Leiden algorithm [35]: This algorithm can be seen
approximated with a simulation, which therefore reduces
as an improvement of the Louvain algorithm. The Leiden
the algorithm complexity.
algorithm also takes advantage of the idea of speeding up
• OpenOrd [23]: This algorithm aims to better distinguish
the local movement of nodes and the idea of moving nodes
clusters via a simulated annealing type schedule. Long
to random neighbours. It consists of three phases: (1) local
edges are cut to allow clusters to separate.
movement of nodes, (2) refinement of the partition and (3)
2.4 Analysis and Additional visualisation their respective node sizes, and the second one describes the
In addition to the Gephi [6] analysis of the graph, a web-based directed edges that connect these nodes.
platform has been developed with the aim of performing interac- As it has been aforementioned, a simplification of this data is
tions on the graph previously built. This web, based on Node.js designed, grouping followers in clusters with the same pattern
[13], provides navigation options such as pan or scroll through with respect to what politicians they follow. This simplification
the data in order to drill-down and access details of each part of leads to a nodes table with 124.377 entries, and an edges table with
the graph. Selecting and zooming can be used to facilitate quick 898.309 rows. It should be noted that this simplification does not
and interactive exploration of data in order to see connections imply removal of information, as the individual data of each user
between nodes and communities. Smooth zooming is used in is not relevant for this paper’s purposes. However, this data has
order to explore a region of interest. many negligible nodes due to their small size, so a filter has been
Furthermore, it could be possible to visualise statistics related applied, removing those whose node size is less or equal than 100.
to a set of connected nodes by selecting a particular node. This After that, the size of the nodes and edges tables are considerably
analysis is based on the following. reduced to 2.863 and 12.291, respectively. These removed clusters
As it has been stated in Subsection 2.1, the initial users associ- are representative of such a small portion of individuals (less
ated to the followers of the studied politicians have been analysed than 100 users in a population of over 6 million), that the vast
also with respect to which political party or parties are they in- visualisation improvement makes up for the information loss.
terested in, obtaining a political profile for more than 4 million This process allows a better visualisation, as such a large number
of users for the ten selected parties. Furthermore, an automatic of nodes and edges would lead to a hard to interpret graph with
daily process has been designed to download every tweet that insignificant noise.
mentions the leaders and official accounts of the main six politi- On the other hand, the parallel process to obtain political
cal parties (PSOE, PP, VOX, Unidas Podemos, Ciudadanos, Más profiles has been also applied. With this process, 4.567.607 ac-
País). Once this information is processed as well, it is possible counts have been identified with real interest in at least one of
to merge both datasets in order to obtain valuable information, the studied political parties. In Table 2, a summary of these users
being able to analyse, in a daily basis, the profile of the authors is provided, with a list of the total number of individuals that
of every downloaded message. have been identified as interested in each political party.
Thus, it is possible to analyse, for each party, what are the
political profiles that speak the most about it. Inversely, the most Table 2: Number of users identified as interested in each
common political targets of the messages generated by a given political party
profile can be obtained as well.
This information is generated every day, so it is possible to Political party Number of interested users
obtain temporal series associated to each political party with PSOE 843.226
respect to both aspects: the profile of those who speak about PP 553.048
them, and the profile of those who they speak about. This eases VOX 300.569
the visualisation of such information, being able to get a global Unidas Podemos 2.332.736
image and evolution of the opinion involving each party in both Ciudadanos 885.340
directions. Más País 521.249
ERC 671.540
3 APPLICATION DEVELOPMENT Junts Per Catalunya 203.204
All the experiments have been performed on an Comet Lake PNV 16.908
i7-10710u CPU 1100 MHz, with 32 GB RAM and a GEFORCE EH Bildu 41.466
GTX1650 MAX-Q graphic card, running Ubuntu 18.04.3 LTS.
Note that the sum of these values is greater than the previous
3.1 Data Collection and Processing number of analysed accounts, as many of them show interest
The data extraction process using the Twitter API is conducted in more than one party simultaneously. Particularly, the three
with the statistical programming language R, which provides a most common combinations are PSOE & Unidas Podemos, Unidas
package developed to apply the different functionalities of the Podemos & Más País and ERC & Junts Per Catalunya with 310.527,
Twitter API (rtweet) [20], after obtaining a mandatory API key. 187.417 and 161.408 occurrences, respectively. These pairs are
Twitter API is free, although it has volume restrictions by period coherent, given that the ones that compose the first two pairs are
of time, so once this limit is reached, the download is stopped for the most voted national left wing political parties, and that ERC
about 15 minutes. However, there are no restrictions concerning and Junts Per Catalunya are both in favour of the independence
the total downloadable volume, so it is possible to obtain the of Catalonia.
required information. Additionally, as it has been previously explained, tweets men-
The 96 political accounts distributed in ten parties as stated tioning the accounts of the six political national parties and their
in Table 1, have been studied by extracting all the users that fol- respective leaders are downloaded daily. This process has been
low each of the these accounts. As a result, a total of 23.305.485 started in January of 2019, and up to November of 2019, the total
connections (edges of the graph), corresponding to 6.241.682 of downloaded messages is over 15 million tweets, with about
unique users (nodes of the graph) have been obtained. This in- 50.000 tweets downloaded every day in average.
formation has been stored in a MongoDB database, as well as
every additional information generated from this initial dataset. 3.2 Experimental Design for the Application
Consequently, two tables have been generated and stored in the First of all, experiments have been carried out in order to compute
database, where the first one describes the different nodes with several centrality measures of the initial graph. In particular, the
computed metrics were the average degree, the graph density and Table 4: Results of the community detection algorithms
the number of connected components. Secondly, the performance
of the two community detection algorithms: Louvain [9] and Algorithm Nº clusters Modularity
Leiden [35] was tested. The method with the best behaviour in Louvain 9 0.990
terms of modularity was selected. The selected configuration for Leiden 9 0.996
these two methods was resolution equal to 0.01, 10 iterations and
random seed for initializing the algorithms.
In order to improve the representation of the graph with its same than the number of communities obtained by the tested
corresponding communities, an adjustment in colour and size algorithms.
was done as previously explained in Subsection 2.3. Regarding In particular, the characteristics of the obtained clusters can
the configuration, the size values were ranging between 10 and be summarized as follows. On the one hand, a significant cluster
200, where 10 corresponds to the minimum size and 200 to the that groups together the regionalist parties from Catalonia with
maximum one. their followers was detected. On the other hand, the algorithm
Following these tasks, ForceAtlas2 [18] and OpenOrd [23] identified two different clusters composed by regionalist parties
layout algorithms were consecutively combined to customize from Basque Country and followers. The VOX party accounts
the network in order to improve its visualisation. The initial and its followers, were represented in a fourth community, which
parameters for ForceAtlas2 were shown in Table 3. On the other appears clearly separated from the other groups. In another clus-
hand, the OpenOrd algorithm was set with an edge cut of 0.95 ter, some political accounts from Más País and Unidas Podemos
(a higher cutting means a more clustered result), a number of leaders were classified. However, followers and other political
iterations of 250 and random seed. accounts of these latter parties were also identified in another
After this, a set of centrality measures was computed for giving community, closer to regionalist parties. Regarding the national
insights about the connections and relationships between commu- parties, the main leaders and political accounts of Ciudadanos
nities: in-degree, out-degree, Betweeness Centrality, Closeness and PP were group together. Additionally, the current socialist
Centrality and Eigenvector Centrality [19]. president and other accounts from PSOE composed the eight
cluster. Finally, the algorithm differentiated in a separate and cen-
Table 3: ForceAtlas2 initial configuration tral community, the Spanish ex-president from the PP party as
well as the PP and PSOE official Twitter accounts and significant
Parameter Value politicians from these two parties.
Threads number 8 With respect to the centrality metrics, four different options
Tolerance 0.9 have been considered: in-degree (number of adjacent incoming
Approximate repulsion yes edges to each node), Closeness (steps required to access every
Approximation 1.2 other node), Betweeness (based on the number of shortest paths
Scaling 1.5 between nodes that pass through a particular node) and Eigen-
Stronger gravity no vector Centrality (connection to well-connected nodes), These
Gravity 1.0 measures have been obtained for every political node that has at
Dissuade hubs no least a connection with two clusters. In order to aggregate that
LinLog mode no information, given the heterogeneous scale, a ranking has been
Prevent overlap yes generated for each node, with greater score to the ones with the
Edge weight influence 1.0 better values. The sum of such values defined the final rank of
these nodes. The most connected nodes are the ones shown in
Table 5.
Finally, the parameters of both the resulting graph and the
layout were saved in a JSON file with the aim of exporting it to the
web application based on Node.js [13]. This front-end constitutes Table 5: Best political nodes with respect to the centrality
a web platform for allowing different actions on the graph like metrics
visualisation, zoom navigation and additional data analysis of
the relevant nodes by means of the interaction with them. Rank Account Party Score
1 @Pablo_Iglesias_ Unidas Podemos 352
3.3 Results 2 @Albert_Rivera Ciudadanos 345
Regarding the metrics that calculate the network properties as 3 @ahorapodemos Unidas Podemos 342
a whole, the obtained results were the following. For the graph 4 @ManuelaCarmena Más País 335
density, the value was equal to 0.003. Closely related to the density 5 @sanchezcastejon PSOE 335
of the graph is the average degree. The value obtained for this 6 @agarzon Unidas Podemos 334
metric was 3.24, i.e., the average number of edges connected to 7 @marianorajoy PP 332
a node. Finally, the number of strongly connected components 8 @KRLS Junts Per Catalunya 326
were nine, being each component a maximal strongly connected 9 @ierrejon Más País 322
subgraph. 10 @PSOE PSOE 314
Table 4 shows the results achieved by the community detection 11 @gabrielrufian ERC 313
algorithms. Both methods obtain nine significant communities ... ... ... ...
inside the network, while the Leiden algorithm outperforms the 20 @vox_es VOX 270
Louvain method in a 0.6%. These results are consistent with the 33 @ArnaldoOtegi EH Bildu 213
fact that the number of strongly connected components is the 47 @eajpnv PNV 167
Figure 2: From left to right evolution of graph visualisation: initial and in-progress layout options
These results show that the nodes with the best centrality speaking about it, and what party are its followers talking about
trade-off are coped by accounts from Unidas Podemos (1st , 3rd , as well.
6th ), while parties like VOX, EH Bildu and PNV get their first All the obtained visual representations have been merged into
account in that ranking in the positions 20st , 33th and 47th , re- one in a website designed to that use. Thus, a graph showing the
spectively. political nodes, the clusters, their connections and the generated
Additionally, the visualisation of the graph by using ForceAt- communities, is presented interactively. Additionally, once a po-
las2 [18] and OpenOrd [23] obtained a set of representations litical node is selected, time series are shown with respect to the
depicted in Figure 2. The final layout was selected for being opinion from and to that party.
saved and represented on the web platform. It is possible to draw important conclusions from this analysis,
In Figure 3, a screenshot of the most dense part of the graph such as the political communities where most accounts from the
from the aforementioned web tool is presented. It is possible to same party are set together despite the system not knowing the
spot a group of political nodes corresponding to parties in favour corresponding political party, as well as the closeness between
of the independence of Catalonia (ERC, Junts Per Catalunya), as parties that share some common ideology. Additionally, opinion
well as the national party VOX isolated from the rest. The other analysis shows the existence of parties whose followers generate
national parties are all together in the centre of the graph, al- a great share of the content directed to this particular party.
though with more closeness between those with more ideological Particularly, the resulting graph shows that the Catalonian
similarities. parties (ERC, Junts Per Catalunya) are in the same cluster taking
Figure 4 presents the additional graphics that show up once into account their common ideology regarding Catalonia inde-
a political node is selected in the interactive graph. The first pendence. VOX is also isolated from the others, showing their
one provides information about the messages written by the auto-connection and lack of interaction with the others. The rest
followers of the selected party. The second one represents the of national parties (PSOE, PP, Unidas Podemos, Ciudadanos, Más
share of opinion directed to that party with respect to the political País) are together, but with visible closeness of those with more
profile of their authors. ideological similarities.
Future work leads to enrich the additional information pro-
4 CONCLUSIONS vided about each party interactively in the graph along with the
opinion time series, using the Twitter profile information of each
A methodology has been proposed to obtain valuable political
analysed user. Additionally, new lines of analysis of politics in
information from Twitter. Particularly, the Spanish scenario has
Twitter are to be considered, which can complement the already
been considered, taking into account the second general election
developed ones.
of 2019, motivated by the changing political situation, where in
five years the system went from bipartisan to multi-party with
six national parties and several regionalist ones. REFERENCES
This analysis is twofold. Firstly, several political accounts from [1] Apoorv Agarwal, Boyi Xie, Ilia Vovsha, Owen Rambow, and Rebecca Passon-
neau. 2011. Sentiment Analysis of Twitter Data. In Proceedings of the Workshop
the ten main political parties have been selected, and their fol- on Languages in Social Media (LSM ’11). Association for Computational Lin-
lowers have been obtained through the official Twitter API. This guistics, Stroudsburg, PA, USA, 30–38. http://dl.acm.org/citation.cfm?id=
2021109.2021114
information, after the necessary processing, is treated as node [2] Akshay Amolik, Niketan Jivane, Mahavir Bhandari, and M. Venkatesan. 2016.
and edge data, and as a consequence, a proper graph analysis and Twitter Sentiment Analysis of Movie Reviews using Machine Learning Tech-
visualisation is provided. Secondly, tweets mentioning the official niques. International Journal of Engineering and Technology 7 (01 2016), 2038–
2044.
accounts of the main six national parties and their respective [3] Michael J. Bannister, David Eppstein, Michael T. Goodrich, and Lowell Trott.
leaders are downloaded in a daily basis. These tweets are charac- 2013. Force-Directed Graph Drawing Using Social Gravity and Scaling. In
terised with respect to the political profile of their authors with Graph Drawing, Walter Didimo and Maurizio Patrignani (Eds.). Springer Berlin
Heidelberg, Berlin, Heidelberg, 414–425.
the data generated by the followers information. Consequently, [4] Pablo Barberá. 2015. Birds of the Same Feather Tweet Together: Bayesian Ideal
it is possible to determine for each party, whose followers are Point Estimation Using Twitter Data. Political Analysis 23, 1 (2015), 76–91.
Figure 3: Screenshot of the central part of the graph in the web platform
[7] Gema Bello-Orgaz, Julio Hernandez-Castro, and David Camacho. 2017. Detect-
ing discussion communities on vaccination in twitter. Future Generation Com-
puter Systems 66 (2017), 125 – 136. https://doi.org/10.1016/j.future.2016.06.032
[8] David R. Bild, Yue Liu, Robert P. Dick, Z. Morley Mao, and Dan S. Wallach.
2015. Aggregate Characterization of User Behavior in Twitter and Analysis of
the Retweet Graph. ACM Trans. Internet Technol. 15, 1, Article 4 (March 2015),
24 pages. https://doi.org/10.1145/2700060
[9] Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne
Lefebvre. 2008. Fast unfolding of communities in large networks. Journal of
Statistical Mechanics: Theory and Experiment 2008, 10 (2008), P10008. http:
//stacks.iop.org/1742-5468/2008/i=10/a=P10008
[10] Johan Bollen, Alberto Pepe, and Huina Mao. 2009. Modeling public mood
and emotion: Twitter sentiment and socio-economic phenomena. CoRR
abs/0911.1583 (2009). arXiv:0911.1583 http://arxiv.org/abs/0911.1583
[11] Meeyoung Cha, Hamed Haddadi, Fabrício Benevenuto, and Krishna P. Gum-
madi. 2010. Measuring User Influence in Twitter: The Million Follower Fallacy.
AAAI Conference on Weblogs and Social Media 14.
[12] Alfonso Crisci, Valentina Grasso, Paolo Nesi, Gianni Pantaleo, Irene Paoli, and
Imad Zaza. 2018. Predicting TV programme audience by using twitter based
metrics. Multimedia Tools and Applications 77, 10 (01 May 2018), 12203–12232.
https://doi.org/10.1007/s11042-017-4880-x
[13] Ryan Lienhart Dahl. 2009 (accessed 30-November-2019). Node.js Foundation.
https://nodejs.org/.
[14] Sahar A. El Rahman, Feddah A. AlOtaibi, and Wejdan. A. AlShehri. 2019. Sen-
timent Analysis of Twitter Data. In 2019 International Conference on Computer
and Information Sciences (ICCIS). 1–4. https://doi.org/10.1109/ICCISci.2019.
8716464
[15] Tarek Elghazaly, Amal Mahmud, and Hesham Hefny. 2016. Political Sentiment
Analysis Using Twitter Data. 1–5. https://doi.org/10.1145/2896387.2896396
[16] Ratab Gull, Umar Shoaib, Saba Rasheed, Washma Abid, and Beenish Zahoor.
Figure 4: Graph after the selection of a political node (top), 2016. Pre Processing of Twitter’s Data for Opinion Mining in Political Context.
Procedia Computer Science 96 (2016), 1560 – 1570. https://doi.org/10.1016/j.
with visualisation of the opinion from (middle) and to procs.2016.08.203 Knowledge-Based and Intelligent Information Engineering
(bottom) the selected party (PSOE) Systems: Proceedings of the 20th International Conference KES-2016.
[17] Derek L. Hansen, Ben Shneiderman, Marc A. Smith, and Itai Himelboim.
2020. Chapter 1 - Introduction to social media and social networks. In
Analyzing Social Media Networks with NodeXL (Second Edition) (second edition
ed.), Derek L. Hansen, Ben Shneiderman, Marc A. Smith, and Itai Himelboim
(Eds.). Morgan Kaufmann, 3 – 10. https://doi.org/10.1016/B978-0-12-817756-3.
https://doi.org/10.1093/pan/mpu011 00001-7
[5] Pablo Barberá and Thomas Zeitzoff. 2017. The New Public Address Sys- [18] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bas-
tem: Why Do World Leaders Adopt Social Media? International Stud- tian. 2014. ForceAtlas2, a Continuous Graph Layout Algorithm for Handy
ies Quarterly 62, 1 (10 2017), 121–130. https://doi.org/10.1093/isq/sqx047 Network Visualization Designed for the Gephi Software. PLOS ONE 9, 6 (06
arXiv:http://oup.prod.sis.lan/isq/article-pdf/62/1/121/24431285/sqx047.pdf 2014), 1–12. https://doi.org/10.1371/journal.pone.0098679
[6] Mathieu Bastian, Sebastien Heymann, and Mathieu Jacomy. 2009. Gephi: An [19] Yuntao Jia, Victor Lu, Jared Hoberock, Michael Garland, and John C. Hart.
Open Source Software for Exploring and Manipulating Networks. http: 2012. Chapter 2 - Edge v. Node Parallelism for Graph Centrality Metrics. In
//www.aaai.org/ocs/index.php/ICWSM/09/paper/view/154
GPU Computing Gems Jade Edition, Wen mei W. Hwu (Ed.). Morgan Kaufmann,
Boston, 15 – 28. https://doi.org/10.1016/B978-0-12-385963-1.00002-2
[20] Michael W. Kearney. 2019. rtweet: Collecting Twitter Data. https://cran.
r-project.org/package=rtweet R package version 0.6.9.
[21] Vishal. A. Kharde and Sheetal. Sonawane. 2016. Sentiment Analysis of Twitter
Data : A Survey of Techniques. CoRR abs/1601.06971 (2016). arXiv:1601.06971
http://arxiv.org/abs/1601.06971
[22] Efthymios Kouloumpis, Theresa Wilson, and Johanna Moore. 2011. Twitter
Sentiment Analysis: The Good the Bad and the OMG! ICWSM.
[23] Shawn Martin, W. Michael Brown, Richard Klavans, and Kevin W. Boyack.
2011. OpenOrd: an open-source toolbox for large graph layout. In Visualization
and Data Analysis 2011, Pak Chung Wong, Jinah Park, Ming C. Hao, Chaomei
Chen, Katy Börner, David L. Kao, and Jonathan C. Roberts (Eds.), Vol. 7868.
International Society for Optics and Photonics, SPIE, 45 – 55. https://doi.org/
10.1117/12.871402
[24] Mark E. J. Newman and Michelle Girvan. 2004. Finding and evaluating com-
munity structure in networks. Phys. Rev. E 69 (Feb 2004), 026113. Issue 2.
https://doi.org/10.1103/PhysRevE.69.026113
[25] Tahir M. Nisar and Man Yeung. 2018. Twitter as a tool for forecasting stock
market movements: A short-window event study. The Journal of Finance and
Data Science 4, 2 (2018), 101 – 119. https://doi.org/10.1016/j.jfds.2017.11.002
[26] Fabiola S. F. Pereira, Sandra de Amo, and João Gama. 2016. Evolving Cen-
tralities in Temporal Graphs: A Twitter Network Analysis. In 2016 17th IEEE
International Conference on Mobile Data Management (MDM), Vol. 2. 43–48.
https://doi.org/10.1109/MDM.2016.88
[27] Ioannis Pitas. 2016. Graph-Based Social Media Analysis. CRC Press. https:
//books.google.es/books?id=BvYYCwAAQBAJ
[28] John Price, Neil Farrington, and Lee Hall. 2013. Changing the game? The
impact of Twitter on relationships between football clubs, supporters and the
sports media. Soccer & Society 14, 4 (2013), 446–461. https://doi.org/10.1080/
14660970.2013.810431 arXiv:https://doi.org/10.1080/14660970.2013.810431
[29] Hassan Saif, Yulan He, and Harith Alani. 2012. Semantic Sentiment Analysis
of Twitter. In The Semantic Web – ISWC 2012, Philippe Cudré-Mauroux, Jeff
Heflin, Evren Sirin, Tania Tudorache, Jérôme Euzenat, Manfred Hauswirth,
Josiane Xavier Parreira, Jim Hendler, Guus Schreiber, Abraham Bernstein, and
Eva Blomqvist (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 508–524.
[30] Wernard Schmit and Sander Wubben. 2015. Predicting Ratings for New
Movie Releases from Twitter Content. 122–126. https://doi.org/10.18653/v1/
W15-2917
[31] Shiladitya Sinha, Chris Dyer, Kevin Gimpel, and Noah Smith. 2013. Predicting
the NFL using Twitter. Proc. ECML/PKDD Workshop on Machine Learning and
Data Mining for Sports Analytics (10 2013).
[32] Michael Steketee, Atsushi Miyaoka, and Maura Spiegelman. 2015. Social
Network Analysis. In International Encyclopedia of the Social Behavioral
Sciences (Second Edition) (second edition ed.), James D. Wright (Ed.). Elsevier,
Oxford, 461 – 467. https://doi.org/10.1016/B978-0-08-097086-8.10563-X
[33] Narges Tabari, Piyusha Biswas, Bhanu Praneeth, Armin Seyeditabari, Mirsad
Hadzikadic, and Wlodek Zadrozny. 2018. Causality Analysis of Twitter Senti-
ments and Stock Market Returns. In Proceedings of the First Workshop on Eco-
nomics and Natural Language Processing. Association for Computational Lin-
guistics, Melbourne, Australia, 11–19. https://doi.org/10.18653/v1/W18-3102
[34] Marijn ten Thij, Tanneke Ouboter, Daniël Worm, Nelly Litvak, Hans Berg, and
Sandjai Bhulai. 2015. Modelling of Trends in Twitter Using Retweet Graph
Dynamics. https://doi.org/10.1007/978-3-319-13123-8_11
[35] Vincent A. Traag, Ludo Waltman, and Nees Jan van Eck. 2019. From Louvain
to Leiden: guaranteeing well-connected communities. Scientific Reports 9,
5233 (2019).
[36] Jo Williams, Susan J Chinn, and James Suleiman. 2014. The value of Twitter
for sports fans. Journal of Direct, Data and Digital Marketing Practice 16, 1 (01
Jul 2014), 36–50. https://doi.org/10.1057/dddmp.2014.36
[37] Yuto Yamaguchi, Tsubasa Takahashi, Toshiyuki Amagasa, and Hiroyuki Kita-
gawa. 2010. TURank: Twitter User Ranking Based on User-Tweet Graph
Analysis. In Web Information Systems Engineering – WISE 2010, Lei Chen,
Peter Triantafillou, and Torsten Suel (Eds.). Springer Berlin Heidelberg, Berlin,
Heidelberg, 240–253.
[38] Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu. 2014. Social media
mining: an introduction. Cambridge University Press.