Online Collaborative Prediction of Regional Vote Results Vincent - - PowerPoint PPT Presentation

Online Collaborative Prediction

• f Regional Vote Results

Vincent Etter, Emtiyaz Khan, Mattias Grossglauser, Patrick Thiran DSAA — October 17, 2016 — Montréal, Canada

Data Opportunity

• Many countries adopt open government initiatives
• Several datasets published
• Demographics
• State affairs
• Votes and elections
• Unique opportunity
• Get a better understanding
• Build tools useful to others

2

Voting Data

• News agencies, political parties, and polling institutes are all

interested in understanding voting behaviors

• Will the next vote pass easily?
• What makes two regions vote similarly?
• Where should we focus our efforts?

3

Dataset

• Vote results from Switzerland
• Issue votes between 1981 and 2014
• Outcome (% of “yes”) at the municipality level
• 281 votes
• 13 features: voting recommendation of the main parties
• 2352 regions
• 25 features: languages spoken, demographics, etc.

4

Data available at http://vincent.etter.io/dsaa16

Similarities Between Results

5

Online Predictions

• On the day of the vote, regional results are released in

sequence

• Use published results to predict others
• … and refine the prediction as more results are published?

6

Our Approach

• Use a matrix-factorization model to capture the bi-clustering
• Add region and vote features
• Reduce the cold-start problem
• More interpretable
• Build the model incrementally to assess the effect of each

component

7

Our Model

8

ydn = zdn + ✏

zdn = µn + fn(xd) + fd(wn) + vT

d un

bias regression

• n region

regression

• n vote

matrix factorization

Our Models

9

zdn = µn + fn(xd) + fd(wn) + vT

d un

γT

d wn

zdn = µn +

LIN(v)

γT

d wn

zdn = µn + + βT

n xd

LIN(r) + LIN(v)

zdn = µn +

vT

d un

+

MF

βT

n xd

zdn = µn + vT

d un

+

MF + LIN(r)

zdn = µn + vT

d un

+ GP(xd)

MF + GP(r)

zdn = µn +

vT

d un

+

GP(xd)

γT

d wn

+

MF + GP(r) + LIN(v) LIN(r) zdn = µn +

βT

n xd

λβ, λγ, λu, λv θ, σs, λγ

Performance Evaluation

• Last 50 votes as test data
• Simulate 500 random reveal order
• Last 10% of regions as test regions
• Observe increasing number of regions
• Predict result of test regions

10

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] LIN(r)

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] LIN(r) MF

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] LIN(r) MF MF + LIN(r)

Results

11

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] MF + LIN(r)

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] MF + LIN(r) M F + G P ( r )

Bayesian VS Non-Bayesian

12

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] LIN(v) M F + G P ( r )

100 101 102 103

Number of observed regions

5 6 7 8 9 10 11 12 13

RMSE on the last 10 % of regions [%] LIN(v) M F + G P ( r ) MF + GP(r) + LIN(v)

Final Model

13

Interpretation

14

0.0 0.2 0.4 0.6 0.8 1.0

Relative importance

Speaks Italian Speaks German Population Jobs Speaks Romansh Population density Speaks French Age 65+ Social aid Foreigners Election PEV Election FDP Election GL Election Greens Age 0-19 Election SP Elevation Election PST Election other right Age 20-64 Election SVP Election BDP Election CVP y x

Röstigraben

Summary

• Individual models have different strengths
• Vote features regression for cold start
• Region features and bi-clustering when more observations
• Bayesian methods are useful
• Proper hyperparameters setting
• Accurate and interpretable results

15

Thank you!

Code and data available at http://vincent.etter.io/dsaa16 Any questions?

16