=Paper=
{{Paper
|id=Vol-1751/AICS_2016_paper_42
|storemode=property
|title=Dimensionality Reduction and Visualisation Tools for Voting Records
|pdfUrl=https://ceur-ws.org/Vol-1751/AICS_2016_paper_42.pdf
|volume=Vol-1751
|authors=Igor Brigadir,Derek Greene,James P. Cross,Padraig Cunningham
|dblpUrl=https://dblp.org/rec/conf/aics/BrigadirGCC16
}}
==Dimensionality Reduction and Visualisation Tools for Voting Records==
https://ceur-ws.org/Vol-1751/AICS_2016_paper_42.pdf
Dimensionality Reduction and Visualisation
Tools for Voting Records
Igor Brigadir1 , Derek Greene1 , James P. Cross2 , Pádraig Cunningham1
1
Insight Centre for Data Analytics, University College Dublin, Ireland
{igor.brigadir,derek.greene,padraig.cunningham}@insight-centre.org
2
School of Politics & International Relations, University College Dublin, Ireland
james.cross@ucd.ie
Abstract. Recorded votes in legislative bodies are an important source
of data for political scientists. Voting records can be used to describe par-
liamentary processes, identify ideological divides between members and
reveal the strength of party cohesion. We explore the problem of working
with vote data using popular dimensionality reduction techniques and
cluster validation methods, as an alternative to more traditional scaling
techniques. We present results of dimensionality reduction techniques
applied to votes from the 6th and 7th European Parliaments, covering
activity from 2004 to 2014.
1 Introduction
As a law making body, votes passed in the European Parliament (EP) can have
significant influence on citizens across the European Union. Members of the Eu-
ropean Parliament (MEPs) hold power over the majority of EU legislation, as
well as decisions on budgets and spending. Analysis of votes is not only of inter-
est to researchers, but many interest groups and industries operating within the
EU. To produce insights into legislation and party politics computational ap-
proaches are highly dependant on latent variable models—using point estimates
to make sense of and test theories using voting records [11], speeches [20], party
manifestos [2], expert surveys [17], and more recently social media data [1].
A common theme in these models is the low dimensional reconstruction of
high-dimensional data. Roll call votes, where the vote of each member is recorded
are typically represented as a matrix of legislators with for and against votes,
treating abstentions as missing values. Legislators, in this case MEPs, are repre-
sented as vectors in d dimensions, where each dimension encodes a vote in some
way. Scaling methods are then applied to recover point estimates or produce visu-
alisations. Scaling methods essentially perform dimensionality reduction, trans-
forming data in a high-dimensional space to a space with fewer dimensions—an
n dimensional space Rn where n << d, typically 2 or 3 dimensions are used to
produce interpretable visualisations.
While established methods for inductive scaling of roll call votes exist, there
are many other potential alternatives that remain unexplored. We describe four
such alternatives in Section 3, and formulate a cluster quality-based evaluation
approach, highlighting advantages and drawbacks of each method. We make
the data for 6th and 7th EU parliaments and Python code to reproduce the
approaches on different sets of voting records are made available online3 , so
that political science researchers can explore these alternative approaches when
analysing vote data.
2 Related Work
The NOMINATE [19] family of multidimensional scaling approaches are the
most widely adopted methods for estimating ideal points from roll call data,
and have been applied to European Parliament roll call vote data in [11] where
the main policy dimensions based on this data reveal a dominant left-right di-
mension, as well as evidence of a pro-/anti-Europe dimension. The results of
scaling are often used as features for downstream tasks, such as [17] where ideal
points are used as features in estimating party influence. In [8] roll call votes are
compared to survey responses.
Scaling using text from speeches [20] can be related to the broader task
of dimensionality reduction [16]. Popular scaling methods include Wordfish [15],
and Wordscores [12]. The Wordfish model is applied to EP debates in [20]. While
strong evidence for left-right ideology was not found in the speeches, the results
suggest that legislators express ideology differently through speaking and voting.
In [9] the voting records are combined with text contained in US House and
Senate data, with ideal points estimated for topics such as health, military, and
education.
What all these approaches share is a strong domain-specific focus: scaling
approaches like W-NOMINATE [18] are developed specifically to deal with roll
call votes and not any other kind of data. We propose adapting dimensionality
reduction methods which are not commonly used with roll call data, but have
been previously shown to be effective elsewhere and are widely used across many
other domains.
3 Methods
We cast the problem of roll call vote analysis as a dimensionality reduction
problem. We apply four methods (described below) to roll call voting records
from the 6th and 7th European Parliament, testing alternative ways of encoding
the vote data with different methods.
3.1 Voting in the EU Parliament
MEPs in the parliament are organised into transnational political groups. Group
membership is based on ideological preferences of members from different coun-
3
https://github.com/igorbrigadir/vote2vec
tries, for example: Conservatives in one country will have more policies in com-
mon with conservatives in other countries, than with liberals in their own coun-
try. These groups work together to divide the workload of drafting legislation,
researching policy and other activities. The groups delegate experts to work on
different issues, and agree to follow their instructions on the best voting strategy.
Given this organisation, MEPs have strong incentives to follow the voting pat-
terns of their group [10]. The groups and their broad ideologies are summarised
in Table 1. MEPs do not always follow group voting decisions, but have strong
incentives to do so, as the groups control allocation of resources and committee
positions.
3.2 Encoding Vote Data
The EP plenary votes are publicly available and published regularly4 . Before
applying techniques to roll call votes, we construct the vote matrix X: the high-
dimensional representation of votes—where an entry contains a binary value for
Yes, No, and optionally Abstain, on each vote by an individual MEP.
A small example representing this encoding
for two roll call votes for three different MEPs
Vote 1 (Yes) is shown in Figure 1.
Other potential encodings, given vote
Vote 1 (No) metadata and method choice are possible: a
count matrix is produced by merging votes us-
Vote 1 (Abstain) ing title similarity, or policy area or committee.
Detailed vote meta data is available for the 6th
Vote 2 (Yes) parliament5 from [10], but is incomplete for the
7th parliament. Results are reported for vote
Vote 2 (No) encoding using individual votes.
MEPs who switch groups [6] during the
Vote 2 (Abstain) term present a data consistency challenge for
roll call analysis using our proposed evaluation
MEP 1 MEP 2 MEP 3
measure. MEPs who follow group voting proce-
Fig. 1: Example vote matrix: dure of one group for a period of the term, and
MEPs 1 and 3 voted No on then switch will be correctly clustered with the
Vote 1, and abstained on Vote group most similar to them, but mislabelled
2. MEP 2 Voted Yes for both. during evaluation, as voting records remain,
while group affiliation can change.
Every effort has been made to correct in-
consistencies with data such as removing du-
plicate vote records and matching roll call records with MEP profiles to ensure
MEPs represent the correct group at the time of the vote, but some inconsisten-
cies may remain.
4
http://www.europarl.europa.eu/plenary/en/votes.html
5
http://personal.lse.ac.uk/hix/HixNouryRolandEPdata.htm
Name Abbreviation Seats Ideology
7th Term 2009–2014
European People’s Party (Christian Democrats) EPP 274 Conservative
Progressive Alliance of Socialists and Democrats S&D 195 Socialist
Alliance of Liberals and Democrats for Europe ALDE 85 Liberal
European Conservatives and Reformists Group ECR 56 Eurosceptic
Greens / European Free Alliance G-EFA 58 Green
Group of the European United Left / Nordic Green Left EUL-NGL 35 Radical Left
Europe of Freedom and Direct Democracy Group EFD 33 Eurosceptic
Non-attached Members NI 30 Various
6th Term 2004-2009
European People’s Party (Christian Democrats) EPP-ED 288 Conservative
Socialist Group in the European Parliament PES 217 Socialist
Alliance of Liberals and Democrats for Europe ALDE 104 Liberal
Union for Europe of the Nations Group UEN 40 Nationalist
Greens / European Free Alliance G/EFA 43 Green
Group of the European United Left / Nordic Green Left EUL/NGL 41 Radical Left
Independence / Democracy Group IND/DEM 22 Eurosceptic
Non-attached Members NI 30 Various
Table 1: Group names, seats, and ideologies for the 6th and 7th parliamentary
terms. Number of seats doesn’t reflect the number of MEPs active over the entire
term, as some retire, or are substituted.
3.3 Dimensionality Reduction
W-NOMINATE: The Weighted Nominal Three-step Estimation approach
[18] is an inductive scaling technique specifically designed for ideal point es-
timation of legislators using roll call data.
While the method is ubiquitous, a number of drawbacks are highlighted in
[3]. Specifically: thresholds that exclude some votes, which results in poorer
discrimination among extremist MEPs, and excluding MEPs with short voting
histories. In the 7th Parliament dataset 5 of 853 MEPs and 460 of 6961 votes
are excluded with the recommended settings. The methods we propose do not
exclude any MEPs or Votes, and do not require setting vote or MEP specific
thresholds, however they do introduce their own method specific parameters and
initialisation strategies that can impact results, and do not solve the problem of
parameter tuning.
PCA: Principle Component Analysis [7] is a commonly used linear dimension
reduction technique. PCA is performed using Singular Value Decomposition on
the vote data matrix. Figures 3 and 4 show the resulting visualisations.
NMF: Given a non-negative matrix X, Non-negative Matrix factorization [14]
approaches find two factor matrices W and H where the product of W and H
approximates X. The dimensions of the factor matrices are significantly lower
than the product. NMF is not commonly used for visualisation, but is a popular
approach for clustering [5] and topic modelling.
t-SNE: t-Stochastic Neighbourhood Embedding is a popular dimensionality
reduction and visualisation technique. Data is usually embedded in two or three
dimensions, creating interpretable visualisations of high dimensional spaces. The
stochastic nature of the process can sometimes produce visualisations that are
drastically different, or contain structure that could be over-interpreted. For
example, in a 2d plot, the x and y coordinates are not reliable values to use as
point estimates in the same way as W-NOMINATE scores are—however, the
clusters produced and relative positions of MEPs can be informative as MEPs
with similar voting patterns will be clustered together.
SGNS with t-SNE: We explore a two step process, where votes and MEPs are
treated as co-occurrences—embedding votes and MEPs into a lower dimensional
space with Stochastic Gradient Descent with Negative Sampling [13] and then
applying t-SNE to further reduce dimensionality down to 2 or 3 for visualisa-
tion. The two step process tends to exaggerate distances between MEPs of the
same group, however this method introduces more parameters and instability,
making qualitative analysis difficult and prone to over interpretation—where
visualisation artefacts can be interpreted as meaningful.
3.4 Evaluating Projections
In order to evaluate the quality of the low dimensional projections of MEPs,
we adopt Within Group Scatter and Between Group Scatter criteria, which
have been widely used for the problem cluster validation [4]. Here we define
our clusters as the parliamentary groups to which MEPs belong. The between
group scatter quantifies differences in voting behaviour between groups, while
within group scatter quantifies how cohesive a group is, or rather, how strongly
party discipline dictates vote behaviour [10].
For group k, the within group scatter is calculated as the within group sum
of squares, or W GSS {k} :
∑ {k}
W GSS {k} = ||Mi − G{k} ||2
i∈Ik
where Gk is the centroid of group k. The between group scatter or BGSS is:
∑
K
BGSS = nk ||G{k} − G||2
k=1
where Gk is the centroid of group k, G is the centroid of all points (representing
MEPs in a 2d space). Small W GSS values indicate tight grouping of points in a
cluster, or strong party discipline in the case of MEPs and votes. Large BGSS
indicates large differences between groups.
4 Results
We now compare the outputs generated by W-NOMINATE and the alternative
methods. Overall, in contrast to W-NOMINATE, the other methods have the
advantage of significantly faster run times, but introduce method specific ini-
tialisations and parameters, which can affect visualisation output. This is most
pronounced in the case of t-SNE with random initialisation, where a cluster of
MEPs may be placed “to the right” or “to the left” of another group depending
on the run. Initialising t-SNE with PCA produces stable arrangements of clus-
ters in a 2d space, but the x and y values of individual MEPs are unsuitable for
use as point estimates.
For W GSS and BGSS we exclude the non attached MEPs, as these are not
members of any political group in the parliament. Ideology in the non-attached
members ranges from communism, to populism, nationalism and neo-nazism.
Figure 2 shows W-NOMINATE estimates that form our baseline: other ap-
proaches are compared to W GSS and BGSS scores derived from these results.
Detailed scores by party group for the parliaments are shown in Tables 2 and 4
below. In Figure 2, the x axis is interpreted as the left/right dimension, with left
wing groups such as the European United Left / Nordic Green Left (EUL/NGL)
placed on the left, and right wing groups such as Europe of Freedom and Democ-
racy (IND/DEM) on the right. The y axis is interpreted as capturing a pro/anti
EU integration dimension, with pro-EU groups assigned estimates close to 1 and
Eurosceptic or anti-EU MEPs assigned point estimates close to -1.
Figures 3 and 4 show an overview of all methods applied to the 6th and
7th parliamentary terms. In contrast to W-NOMINATE, the other methods
have greater within group scatter—exaggerating differences between MEPs in
the same group. While some groups are clustered more appropriately by the
methods we explored, overall W-NOMINATE produces the best clustering of
MEPs.
899 MEPs in 'Vote Space', using W-NOMINATE 712 MEPs in 'Vote Space', using W-NOMINATE
EPP-ED EPP
1.0 PES 1.0 S&D
ALDE ALDE
UEN ECR
G/EFA G/EFA
EUL/NGL EUL/NGL
IND/DEM EFD
NI NI
0.5 0.5
0.0 0.0
0.5 0.5
1.0 1.0
1.0 0.5 0.0 0.5 1.0 1.0 0.5 0.0 0.5 1.0
Fig. 2: W-NOMINATE scales for the 6th (left) and 7th (right) Parliaments.
4.1 6th Term
The 6th term began in 2004, and ended in 2009. In total there are records for
899 MEPs. MEPs sometimes join the parliament at different times, retire, or are
replaced. We include an MEP in a group if they have a record of a vote in the
dataset.
Term 6
wnominate PCA NMF tSNE sgns tsne
Fig. 3: Overview of visualisations built on 6th Term voting records. Points are
in clusters coloured by group.
Group MEPs WNOM. PCA NMF t-SNE SGNS
EPP-ED 340 43.22 139.52 108.88 144.38 101.81
PES 264 11.14 104.53 79.22 74.27 85.12
ALDE 125 1.22 50.96 39.07 32.10 40.65
UEN 51 3.32 16.66 13.73 16.51 12.67
EUL/NGL 48 2.17 16.24 14.39 5.79 11.70
G/EFA 44 1.08 10.47 9.52 2.54 5.71
IND/DEM 27 13.68 9.79 9.14 12.63 8.28
Overall 899 75.84 348.18 273.95 288.23 265.93
Table 2: W GSS: Within Group Scatter, votes from 6th Term. Smaller values indicate
that MEPs in a group are close to other group members in the vote space.
Group MEPs WNOM. PCA NMF t-SNE SGNS
EPP-ED 340 4.88 64.84 124.63 84.77 71.64
PES 264 106.88 45.62 88.63 104.96 6.75
ALDE 125 6.48 3.19 4.61 1.32 16.06
UEN 51 30.13 5.45 10.42 5.13 7.37
EUL/NGL 48 67.56 28.44 49.66 4.32 24.17
G/EFA 44 19.27 34.64 62.29 1.76 32.62
IND/DEM 27 18.92 3.35 2.08 4.57 5.10
Overall 899 254.13 185.53 342.32 206.84 163.71
Table 3: BGSS: Between Group Scatter, using votes from the 6th Term. Larger values
indicate greater separation between clusters of MEPs.
4.2 7th Term
The 7th parliament was elected in 2009 and finished in 2014. Between the 7th and
the 6th parliaments there were a number of changes made to groups, including
new members and affiliation switches with existing MEPs.
Term 7
wnominate PCA NMF tSNE sgns tsne
Fig. 4: Overview of visualisations built on 7th Term voting records. Points are
in clusters coloured by group.
Group MEPs WNOM. PCA NMF t-SNE SGNS
EPP 267 12.18 41.54 36.81 49.82 68.47
S&D 184 6.72 26.75 23.88 27.21 32.23
ALDE 85 0.96 15.55 13.91 12.36 14.33
G/EFA 56 0.70 7.89 7.06 6.88 2.99
ECR 54 3.17 3.80 2.83 2.97 3.20
EUL/NGL 35 2.59 6.45 5.96 5.31 4.60
EFD 31 5.85 6.97 3.91 6.50 4.22
Overall 712 32.17 108.95 94.35 111.06 130.03
Table 4: W GSS: Within Group Scatter, votes from 7th Term. Smaller values indicate
that MEPs in a group are close to other group members in the vote space.
Group MEPs WNOM. PCA NMF t-SNE SGNS
EPP 267 125.35 132.34 281.77 121.42 44.48
S&D 184 1.45 86.05 187.74 76.81 81.13
ALDE 85 22.32 2.64 4.05 2.11 40.16
G/EFA 56 57.96 44.33 92.59 42.30 25.37
ECR 54 39.57 28.56 18.36 29.45 44.77
EUL/NGL 35 23.11 35.09 43.00 31.59 37.77
EFD 31 16.52 20.80 7.84 20.19 25.22
Overall 712 286.29 349.81 635.33 323.86 298.89
Table 5: BGSS: Between Group Scatter, using votes from the 7th Term. Larger values
indicate greater separation between clusters of MEPs.
5 Discussion
While the methods we explore do not out perform the well established and
widely used W-NOMINATE approach using a cluster validation based evalua-
tion, there are a number of useful recommendations we can make when using dif-
ferent methods: NNDSVD initialization strategy for NMF produces most stable
results; PCA initialization for t-SNE can help with stability of results. Even so,
there is still a risk of over interpreting the structure that t-SNE produces. Before
drawing any conclusions from visualisations made with t-SNE, we recommend
paying particular attention to the implementation and parameters, especially
the learning rate used during optimization. The SGNS approach allows most
flexibility with encoding votes, but is the least stable method. The dimensions
themselves from NMF, or t-SNE are not as useful for point estimates compared
to W-NOMINATE, but the relative positions of cluster centroids offer a useful
measure of similarity between groups.
Many techniques are applicable if we treat roll call vote scaling as a dimen-
sionality reduction problem. All methods that aim to project or embed high
dimensional data in a low dimensional space introduce some uncertainty and in-
stability. Uncertainty in point estimates can come from many sources: from data
quality issues and encoding schemes, to parameter and initialization choices,
to visualisation choices. Given these issues, one advantage that the alternative
methods we explored have is their speed and efficiency: multiple runs under
different settings can highlight errors in ideal point estimates more clearly.
In terms of evaluation, expert surveys [17] or coded party manifestos [2] may
offer better benchmarks for differences between groups and MEPs. Producing
annotations and expert surveys is a costly task however, and there are currently
no expert judgements or annotations available for all votes for a full term.
6 Conclusion
We applied several commonly used dimensionality reduction techniques to voting
records in the EU parliament. While all techniques tend to exaggerate distances
between MEPs of the same group, they can perhaps be useful for quantifying
within-party differences, or treating cluster centroids as points—similarities be-
tween groups.
Applying similar methods to speeches and using point estimates derived from
our proposed methods as alternatives in downstream tasks is ongoing, as well as
comparisons of other projection techniques, applied to more recent data covering
the current 8th parliamentary term.
Acknowledgement: This publication has emanated from research conducted
with the support of Science Foundation Ireland (SFI) under Grant Number
SFI/12/RC/2289.
References
1. Barbera, P.: Birds of the same feather tweet together: Bayesian ideal point esti-
mation using twitter data. Political Analysis 23(1), 76–91 (2014)
2. Braun, D., Mikhaylov, S., Schmitt, H.: European parliament election study 2009,
euromanifesto study (2010)
3. Clinton, J., Jackman, S., Rivers, D.: The statistical analysis of roll call data. Amer-
ican Political Science Review 98(2), 355–370 (2003)
4. Desgraupes, B.: Clustering indices. University of Paris Ouest-Lab ModalX 1, 34
(2013)
5. Ding, C.: Nonnegative matrix factorizations for clustering: A survey. Data cluster-
ing: Algorithms and Applications p. 148 (2013)
6. Evans, A.M., Vink, M.P.: Measuring group switching in the european parliament:
Methodology, data and trends (1979-2009). Análise social pp. 92–112 (2012)
7. Fodor, I.K.: A survey of dimension reduction techniques. Tech. rep., Lawrence
Livermore National Lab., CA (US) (2002)
8. Gabel, M., Hix, S.: From preferences to behaviour: Comparing meps survey re-
sponse and roll-call voting behaviour. In: Tenth Biennial Conference of the Euro-
pean Union Studies Association. Citeseer (2007)
9. Gu, Y., Sun, Y., Jiang, N., Wang, B., Chen, T.: Topic-factorized ideal point es-
timation model for legislative voting network. In: Proceedings of the 20th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining. pp.
183–192. KDD ’14, ACM (2014)
10. Hix, S., Noury, A., Roland, G.: Voting patterns and alliance formation in the
european parliament. Philosophical Transactions of the Royal Society of London
B: Biological Sciences 364(1518), 821–831 (2009)
11. Hix, S., Noury, A., Roland, G.: Dimensions of politics in the european parliament.
American Journal of Political Science 50(2), 494–511 (2006)
12. Laver, M., Benoit, K., Garry, J.: Extracting policy positions from political texts
using words as data. The American Political Science Review 97(2), 311–331 (2003)
13. Levy, O., Goldberg, Y.: Neural word embedding as implicit matrix factorization.
In: Advances in Neural Information Processing Systems 27: Annual Conference
on Neural Information Processing Systems 2014, December 8-13 2014, Montreal,
Quebec, Canada. pp. 2177–2185 (2014)
14. Lin, C.J.: Projected gradient methods for nonnegative matrix factorization. Neural
Computation 19(10), 2756–2779 (2007)
15. Lo, J., Proksch, S.O., Slapin, J.B.: Ideological clarity in multiparty competition: A
new measure and test using election manifestos. British Journal of Political Science
FirstView, 1–20 (2014)
16. Lowe, W.: There’s (basically) only one way to do it. Available at SSRN 2318543
(2013)
17. McElroy, G., Benoit, K.: Policy positioning in the european parliament. European
Union Politics 13(1), 150–167 (2012)
18. Poole, K., Lewis, J., Lo, J., Carroll, R.: Scaling roll call votes with wnominate in
r. Journal of Statistical Software 42, 1–21 (2011)
19. Poole, K.T., Rosenthal, H.: Congress: A Political-Economic History of Roll Call
Voting. Oxford University Press (2000)
20. Proksch, S.O., Slapin, J.B.: Position taking in european parliament speeches.
British Journal of Political Science 40(03), 587–611 (2010)