Collaborators: Vincent Etter, Julien Herzen, Emtiyaz Khan, - - PowerPoint PPT Presentation

Matthias Grossglauser, EPFL Collaborators: Vincent Etter, Julien Herzen, Emtiyaz Khan, Victor Kristof, Patrick Thiran EIT ICT Labs Summer School August 2015

1

• Team

goals: modeling large systems

• f

social interactions

• Online

social networks

• Mobility
• Epidemics
• Crowdsourcing
• …
• Democracy

= a rich & complex social system

• Switzerland:

sophisticated political system, direct democracy

• Data:

Open Government initiatives

2

Overv rview iew: In Infor

• rma

mation

• n

an and Ne Network

• rk

Dy Dynam amics cs Gr Grou

• up

• Diversified

party landscape

• Four
• fficial

languages

• smartvote:

available since 2003

• Direct

democracy with frequent issue votes

• n

various subjects

• at

both parliamentary and citizen levels

3

Our lab abor

• ratory

tory: Switzerland erland

4

Switzerland erland: direct ct demo mocr crac acy

• 1:

Smartvote

• 32

questions covering different societal & political themes

• Answers

from candidates and citizens before the election

• f

the Nationalrat in 2011

• ~3’000

candidates (82.4%

• f

all candidates)

• ~220’000

citoyens (9%

• f

active voters)

• 2:

Parliament votes

• ~2’500

votes (2011

• 2013)
• f

the 181 candidates elected in 2011 (with smartvote profile)

• 3:

Federal initiatives (plebiscites)

• Result
• f

245 federal votes (1981

• 2011)

for ~2’400 municipalities

5

Da Data sou

• urce

ces

• Example:
• Chocolate

consumption vs Nobel prizes?

• Visual

detection

• f

relationships and trends

6

Di Dime mension

• nali

ality ty reduct ctio ion: 2-D D is is eas asy

principal direction («component»)

[F. H. Messerli, New Engl J Med 2012]

7

Which ch is is the best «direct ction ion

• f
• f

proj

• jectio

ction» n»?

8

Which ch is is the best «direct ction ion

• f
• f

proj

• jectio

ction» n»?

Principal component

Observations:

• The

principal direction itself is interesting

• Projecting

the data

• nto

the principal directions is

• ften

the best low-dimensional image

• f

the data

• First

example: «chocolate» vs «Nobel awards»

• 2

 1 dimension

• Second

example: elephant in 3-D

• 3

 2 dimensions

• Typical

data: very high dimension

• Smartvote

data: 32 questions (point=candidate)

• Municipality

data: 245 votes (point=municipality)

• Parliament:

data 2500 votes (point=legislator)

• Questions:
• What

are the principal components?

• What

does the data projected into the «ideological space» look like?

9

Hi High-dimen imensi sion

• nal

al data is is har ard

10

11

Sma mart rtvote

• te

datas aset

• smartvote

pre-electoral

• pinions
• f

the 2011 parliamentary elections

• 2,985

candidates (82.4%

• f

all candidates)

• 229,133

citizens (~9%

• f

total turnout)

• Examples
• f

questions:

• “Should

Switzerland embark

• n

negotiations in the next four years to join the EU?”

• “How

much should the public transport budget be?”

• Possible

answers

• strongly

disagree

• disagree
• agree
• strongly

agree

• less
• no

change

• more

12

Sma mart rtvote

• te

datas aset

• What

questions discriminate best the

• pinion
• f

candidates?

• Is

the traditional left/right view meaningful?

• Use

dimensionality reduction to find

• ut
• Use

SVD

• n

the matrix 𝑌 of candidates’ responses

13

Di Discr crimi mina nativ tive questio ions ns

𝑌

• 1:

Center 𝑌

• 2:

Singular Value Decomposition: 𝑌 = 𝑉Σ𝑊𝑈

14

=

left-singular vectors (basis for candidate space) right-singular vectors (basis for ideology space) singular values

Di Dime mension

• nali

ality ty reduct ctio ion

× ×

𝑌′

• Projecting

𝑌 onto first two (right) singular vectors

• 𝑌′ = 𝑌 𝑊{1,2}
• 𝑌′ is

𝐷 × 2

15

=

Di Dime mension

• nali

ality ty reduct ctio ion: n: proj

• ject

ction ion

𝑌

×

Projection

• nto

2-dim “ideology space”

16

Id Ideol

• log
• gical

ical spac ace

Observation:

• A

lot

• f
• verlap

17

Par arty

• v
• verl

rlap ap

Observation:

• Center
• f

a party: median

• f

all candidates

• f

party

• Candidate

𝑦: closer to

• wn

party

• r

another party?

• Plot:

for each party, fraction

• f

candidates closer to another party

1st

st ax

axis

• Seriez-vous

favorable à ce que le droit de vote au niveau communal soit instauré pour les étran angers ers qui vivent en Suisse depuis au moins dix ans et ce, dans toute la Suisse?

• Approuveriez-vous

que la concurr rren ence ce fiscal ale entre les can antons ns soit plus limitée?

• Soutenez-vous

l'initiative populaire qui souhaite que le sal alair aire le plus élevé au sein d'une entr trep epri rise se ne puisse pas être plus de douze fois supérieur au salaire le plus bas versé par la même

• entreprise. (initiative

1:12)?

• Une

initiative populaire souhaite instaurer une cai aisse sse mal alad adie ie unique et publique pour l'assurance de

• base. Êtes-

vous favorable à ce projet?

18

Pri rincip cipal al co comp mpon

• nents

nts

Social questions («égalité»)

2nd

nd ax

axis

• Approuvez-vous

des engagements de soldats armés (pour l'autoprotection) de l'ar armée ée suisse à l'étran anger er dans le cadre de missions de maintien de la paix de l'ONU

• u

de l'OSCE?

• Êtes-vous

en faveur d'un accord de libre-écha échange agricole avec l'UE UE ?

• Êtes-vous

favorable à l'accord sur la libre circula ulatio tion des personnes existant avec l'UE?

• Une

imposition centrale sur les quantités dans la production laitière doit-elle être réinstaurée en Suisse à la place du libr bre mar arché hé laitier?

19

Pri rincip cipal al co comp mpon

• nents

nts

Economics, globalisation («liberté»)

3rd

rd ax

axis

• Seriez-vous

favorables à ce que l'eutha thana nasie sie active directe soit légalement possible par le biais d'un médecin en Suisse?

• Les

couples homose sexu xuel els sous le régime du partenariat enregistrés devraient-ils pouvoir adopter des enfants?

• La

Suisse possède des règles relativement strictes concernant la procré réatio tion médicalement assistée. Celles-ci devrait-elles être assouplies?

• La

consommation ainsi que la possession pour la consommation personnelle de dr drogues es dures et douces doivent-elles être légalisées?

20

Pri rincip cipal al co comp mpon

• nents

nts

Observation:

• Principal

components correspond to clearly interpretable political and ideological dimensions Society, ethics («fraternité»)

• The

density profiles

• f

politicians and citizens are very different

21

Sma mart rtvote

• te:

density ty

• v
• ver

ideol

• log
• gy

spac ace

Candidates Citizens

Observation:

• The

positions

• f

politicians

• n

smartvote are not representative

• f

the general public

• «24

heures»: at least

• ne

party instructs candidates

• n

how to answer questions

• Thought

experiment: could a candidate identify a position in ideology space which is (a) close to many voters, but (b) far from

• ther

candidates?

22

Pol

• litical

ical ma mark rketing ting?

• We

compute

• ptimal

set

• f

answers according to (a) and (b)

• Recomputed

recommendations for the 229133 smartvote users

• Computed

# recommendations by

• ur

«fake» candidate relative to genuine candidates

23

Im Impac act

• f
• f

an an cr craf afted sma mart rtvote

• te

prof

• file

• The

«fake» candidate would be recommended to almost ½ of all smartvote users!

24

Results ts: fa fake vs g genuine ne ca candidate

Observation:

• Candidates

could take «unusual» positions to attract recommendations

• No

indication that this is currently being exploited

25

• Public
• 2,494

since the 2011 elections

26

Par arliament ament vot

• tes

• Citizens

are more spread

• n

the ideological plane, while candidates are more polarized

• This

can be seen in the proportion

• f

the variance captured by each singular vector

27

Pol

• lar

ariz ization tion

Candidates Citizens Parliament

• Is

it possible to detect whether a politician crafted his/her profile, given the way he/she votes

• nce

elected ?

• Approach:

learn a predictor for parliament vote 𝑤 from smartvote answers

• Then

compare predicted and actual vote

28

Do Do pol

• litica

icans ns fl flip-fl flop

• p

af after being elect cted ed?

𝑌

Logit classifier (per candidate)

𝑤 prediction performance

• Logit

classifier predicts ≥ 50%

• f

the votes with ≥ 95% accuracy

29

Sma mart rtvote

• te

respon

• nses

es  predict ct par arliam iament ent vot

• tes

Predictable votes

• Comparison

between votes expected from smartvote responses and actual votes cast in parliament

• Using

votes predicted with accuracy > 95%

30

Opinion

• n

shift fts fr from

• m

ca camp mpai aign gn to par arliament ament?

Observation:

• No

significant flip-flopping effect

31

• Dataset:
• utcome

(% yes)

• f

245 votes since 1981 in 2,398 municipalities

• Dimensionality

reduction highlights linguistic / cultural contrasts

32

Vot

• tin

ing patterns rns in Switzerland erland

“Röstigraben”

Zürich, Bern, Basel Geneva, Lausanne

Explore the data: www.predikon.ch

33

Firs rst pri rincip cipal al co comp mpon

• nent

nt

34

Seco cond pri rincip cipal al co comp mpon

• nent

nt

35

Vot

• tin

ing patterns rns at mu munici cipal al level

www.predikon.ch/eigenmap

Geneva Basel Zürich Bern Lausanne

36

Evol

• lution

tion

french german italian

• Knowing

the result

• f
• ne

municipality in advance (e.g., from polling/survey), can we predict the final result?

• Answer:

Yes, but it depends

• n

which municipality!

37

Predictio ction

• f
• f

n nation

• nal

al results ts

Ebikon (accuracy 95.9% on test set)

x

38

Ebiko kon, the mi mini-Swi witzerla tzerland nd

• How

to predict national

• utcomes

from partial results?

• State
• f

the art: average

• f

municipalities reporting so far

• Might

be biased:

• Small

municipalities might complete earlier

• Different

rules in different cantons/municipalities

• …
• Can

we do better? Learn latent-factor model:

• 𝑍

𝑗𝑜 = 𝑊 𝑗 𝑈𝑉𝑜 + 𝛾𝑗 𝑈𝑦𝑗 + 𝛾𝑜 𝑈𝑦𝑜

• 𝑍

𝑗𝑜:

fraction “yes” in municipality 𝑗 for vote 𝑜

• 𝑗:

municipality, 𝑜: vote, 𝜗𝑗𝑜: noise, 𝑚: latent dim = lengths

• f

𝑊

𝑗, 𝑉𝑜

• Interaction

between municipality and vote through 𝑊

𝑗 𝑈𝑉𝑜

39

Ongoi

• ing

ng wor

• rk:

vot

• te

predictio ction fr from

• m

pa part rtial ial results ts

• Vote

features 𝛾𝑜: environment, social, economic, Europe, …

• Most
• bjective:

party recommendations

• Municipality

𝛾𝑗: size, elevation, density,…

• Easy

to

• btain

from

• fficial

sources & wikipedia

40

Go Going fu furt rther: r: ri rich cher mo models with fe features

41

Predictin cting vot

• te
• u
• utco

comes mes fr from

• m

a fe few mu munici cipali alities ties

Error for average

• ver

sample Error for latent-factor estimate

• ver

sample

• Goals:
• Scientific:

improve methods and tools for data analysis

• Societal:

enrich

• ur

democracy

• Motto:
• Mine

first, ask questions later!

• Explore

yourself:

• predikon.ch,

[COSN2014]

42

Summ mmar ary

Matthias Grossglauser, EPFL Collaborators: Vincent Etter, Julien Herzen, Emtiyaz Khan, Victor Kristof, Patrick Thiran EIT ICT Labs Summer School August 2015

43