infervote.org
Democratizing democracy: a resource for political
engagement
Robert Vogel
Why?
• We are constantly bombarded with political rhetoric that
shape our political views.
• How are we ACTUALLY represented by our elected
officials?
• How does and will our congress vote on topics we care
about?
• Do senator voting records exhibit polarized behavior?
• How can we find misbehaving and polarized senators?
• What action can we take?
Demo
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
Jij =
Votes in common
All votes
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
Jij =
Votes in common
All votes
Ji1 = ~1
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
Jij =
Votes in common
All votes
Ji1 = ~1
Ji2 =
+
~0
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
Jij =
Votes in common
All votes
+
Polarity=
Ji1 = ~1
Ji2 =
+
~0
The Polarity Index
Senator i
Votes
Republican
Senator j
Votes
Votes in common
Jij =
Votes in common
All votes
+
Polarity=
Ji1 = ~1
Ji2 =
+
~0
Clustering Senator Voting with Jaccard distance
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
d
Mitch McConnell (KY)
John McCain (AZ)
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
Elizabeth Warren (MA)
Dianne Feinstein (CA)
d
Mitch McConnell (KY)
John McCain (AZ)
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
Elizabeth Warren (MA)
Dianne Feinstein (CA)
Bernie Sanders (VT)
d
Mitch McConnell (KY)
John McCain (AZ)
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
Elizabeth Warren (MA)
Dianne Feinstein (CA)
Bernie Sanders (VT)
d
Mitch McConnell (KY)
John McCain (AZ)
Rand Paul (KY)
Marco Rubio (FL)
Ted Cruz (TX)
Democrat RepublicanInd
Clustering Senator Voting with Jaccard distance
Elizabeth Warren (MA)
Dianne Feinstein (CA)
Bernie Sanders (VT)
d
Mitch McConnell (KY)
John McCain (AZ)
Rand Paul (KY)
Marco Rubio (FL)
Ted Cruz (TX)
Do votes align with bill sponsors?
Republican Sponsor
• The bill sponsor is the member of congress that
introduces the document for consideration.
Democrat Sponsor
Infer directionality of biochemical reactions using Langevin
dynamics
Robert Vogel
Developed new parameterization of therapeutic drugs
using insight from nonlinear dynamical systems
Voting Distributions and the Simulated Senate
• Sample 5000 experimental senates using parameters from data
• Data exhibit a more diverse distribution then simulation
• Potential next step, use the Ising model to model pairwise interactions
Republican SponsorDemocrat Sponsor
The Jaccard Index and Political Polarity
Jaccard Index for Measuring Polarity
• Jaccard Index measures the number identical votes
between Senator i and Senator j normalized to total votes
• Polarity index is the average Jaccard index between
Senator i and all Senators in party R.
Jij =
|vi  vj|
|vi [ vj|
JiR =
1
NR
X
j2R
Jij
Distribution of polarity index
• If party politics were not a factor, these distributions
would overlap
Jaccard Distance for Senator Clustering
• Jij 1 the more similar Senator i votes to Senator j.
• Hierarchical clustering utilizes a dissimilarity measure.
Standard solution 1 - Jij
dJ (i, j) = 1
|vi  vj|
|vi [ vj|
| {z }
Jij
Votes are strictly partisan
• Fraction of votes along party line, most votes are partisan
Topic Modeling
Topic modeling of legislative summaries
Wordspaceperbill
TopicSpace
Bills
Topics
T S = S’
Y N0
Congress person Topic Probabilities
P
T’ = P’
New bill in topic space
Probability of vote
P
Y
N
0
Prediction
Clustering
Can we make predictions of senator votes from
legislative documents?
Topic 1
Topic 2
Topic 3
Senator 1 Vote
Document 1
Document 2
Document N
Topic M
Can we make predictions of senator votes from
legislative documents?
Topic 1
Topic 2
Topic 3
Senator 1 Vote
Document 1
Document 2
Document N
Topic M
VoteSenator 2
Can we make predictions of senator votes from
legislative documents?
Topic 1
Topic 2
Topic 3
Senator 1 Vote
Document 1
Document 2
Document N
Topic M Senator L Vote
VoteSenator 2
Legislative document reduction to topics
• 2559 legislative summaries
• Constructed 6403 word basis from text by:
• removing stop words (e.g and, that,
this, a)
• removing non-english words
• stemming (e.g. rested equal to rest)
• TF-IDF
• Cosine similarity to group topics
Legislative document reduction to topics
• 2559 legislative summaries
• Constructed 6403 word basis from text by:
• removing stop words (e.g and, that,
this, a)
• removing non-english words
• stemming (e.g. rested equal to rest)
• TF-IDF
• Cosine similarity to group topics
• Result: No structure in bill data,
more data needed!
Documents
Documents
Document dimensionality reduction not sufficient
with PCA
95% of the variablity
corresponds to > 1000
dimensions
A small topic space, represents
a small portion
of the variability
tSNE dimensionality reduction suggests
no structure in bill data
• Each point is a document in the reduced space defined by tSNE
• t-distributed Stochastic Neighborhood Embedding maps points
from a high to a low dimensional space by minimizing the
Kullback-Leibler Divergence (minimize information loss).
The data
Why only choose Bills and Amendments?
• In general, these documents can become law
• Other votes are for approving nominations for office and
resolutions.
• Resolutions can be very diverse as shown below.
Graduate Research: An overview
• Langevin Dynamics to:
• figure out direction in biochemical reactions, and
• testing isolation of a network motif.
• Bifurcation analysis to identify:
• nodes in a network sensitive to therapeutic inhibition
Biochemical Noise
• Flow cytometry measures the
relative quantity of <= 12
biochemical species per cell at
a rate of 20,000 cells per
second.
• Fluorescent molecules are
coupled to antibodies that
specifically bind to a
biochemical species.
• Quantity of molecules is
proportional to fluorescent
signal
−2 −1 0 1 2
−4
−3
−2
−1
0
1
2
3
4
PMA 1
PMA 2
PMA 3
Log2 Normalized pMEK
Log2NormalizedppERK
PMA 1
PMA 2
PMA 3
PMA 1
PMA 2
PMA 3
−2 −1 0 1 2
−4
−3
−2
−1
0
1
2
3
4
PMA : 1
Log2 pMEK
Log2ppERK
PMA : 1
−2 −1 0 1 2
−4
−3
−2
−1
0
1
2
3
4
PMA : 2
Log2 pMEK
Log2ppERK
PMA : 2
−2 −1 0 1 2
−4
−3
−2
−1
0
1
2
3
4
PMA : 3
Log2 pMEK
Log2ppERK
PMA : 3
Fluctuations break symmetry
of average measurements
Variance of Y > X
Y
X𝜉x
𝜉y
O
O
• Fluctuations from source node propagates to target
X
Y𝜉y
𝜉x
O
O
Variance of X > Y
True Model False Model
Fluctuations break symmetry
of average measurements
Variance of Y > X
0.2 0.3 0.4 0.5 0.6
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Cov(pMEK, ppERK)
Variance
True Model
pMEK
ppERK
0.2 0.3 0.4 0.5 0.6
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Cov(pMEK, ppERK)
Variance False Model
pMEK
ppERK
0.2 0.3 0.4 0.5 0.6
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Cov(pMEK, ppERK)
Variance
True Model
pMEK
ppERK
0.2 0.3 0.4 0.5 0.6
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Cov(pMEK, ppERK)
Variance
False Model
pMEK
ppERK
Y
X𝜉x
𝜉y
O
O
• Fluctuations from source node propagates to target
X
Y𝜉y
𝜉x
O
O
Variance of X > Y
True Model False Model
Nonlinear dynamics of biochemical inhibition
Inhibition of biochemical signaling in cells,
a new parameter 𝛼
L c SRC
pMEK MEKi
ppERK
In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
Inhibition of biochemical signaling in cells,
a new parameter 𝛼
L c SRC
pMEK MEKi
ppERK
L c SRC
SRCi
pMEK
ppERK
In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
Inhibition of biochemical signaling in cells,
a new parameter 𝛼
L c SRC
pMEK MEKi
ppERK
L c SRC
SRCi
pMEK
ppERK
In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
Inhibition of biochemical signaling in cells,
a new parameter 𝛼
L c SRC
pMEK MEKi
ppERK
L c SRC
SRCi
pMEK
ppERK
In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
Nonlinear dynamics
of biochemical inhibition in cells
Chemical Species
Chemical Complex
Enzymatic reaction
Enzymatic Inhibition
In preparation
for publication
Nonlinear dynamics
of biochemical inhibition in cells
Chemical Species
Chemical Complex
Enzymatic reaction
Enzymatic Inhibition
L c SRC
pMEK MEKi
ppERK
[MEKi] [MEKi]
In preparation
for publication
Nonlinear dynamics
of biochemical inhibition in cells
Chemical Species
Chemical Complex
Enzymatic reaction
Enzymatic Inhibition
L c SRC
SRCi
pMEK
ppERK
[SRCi] [SRCi]
L c SRC
pMEK MEKi
ppERK
[MEKi] [MEKi]
In preparation
for publication
Finding dysfunctional components
in tumor samples
Single cell measurements find abnormalities in
tumor patient profiles
• Kullback-Leibler divergence measures the dissimilarity of the single cell
distribution of biochemical signaling features between patient and healthy
donor samples.
Sjk =
X
i2HD
DKL (Pj(xk)||Pi(xk))
=
X
i2HD
Pj(xk) log
✓
Pj(xk)
Pi(xk)
◆
• k = Biochemical species
• j = patient id
• i = Healthy donor

Final demo

  • 1.
    infervote.org Democratizing democracy: aresource for political engagement Robert Vogel
  • 2.
    Why? • We areconstantly bombarded with political rhetoric that shape our political views. • How are we ACTUALLY represented by our elected officials? • How does and will our congress vote on topics we care about? • Do senator voting records exhibit polarized behavior? • How can we find misbehaving and polarized senators? • What action can we take?
  • 3.
  • 4.
    The Polarity Index Senatori Votes Republican Senator j Votes
  • 5.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common
  • 6.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common Jij = Votes in common All votes
  • 7.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common Jij = Votes in common All votes Ji1 = ~1
  • 8.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common Jij = Votes in common All votes Ji1 = ~1 Ji2 = + ~0
  • 9.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common Jij = Votes in common All votes + Polarity= Ji1 = ~1 Ji2 = + ~0
  • 10.
    The Polarity Index Senatori Votes Republican Senator j Votes Votes in common Jij = Votes in common All votes + Polarity= Ji1 = ~1 Ji2 = + ~0
  • 11.
    Clustering Senator Votingwith Jaccard distance
  • 12.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance
  • 13.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance d Mitch McConnell (KY) John McCain (AZ)
  • 14.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance Elizabeth Warren (MA) Dianne Feinstein (CA) d Mitch McConnell (KY) John McCain (AZ)
  • 15.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance Elizabeth Warren (MA) Dianne Feinstein (CA) Bernie Sanders (VT) d Mitch McConnell (KY) John McCain (AZ)
  • 16.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance Elizabeth Warren (MA) Dianne Feinstein (CA) Bernie Sanders (VT) d Mitch McConnell (KY) John McCain (AZ) Rand Paul (KY) Marco Rubio (FL) Ted Cruz (TX)
  • 17.
    Democrat RepublicanInd Clustering SenatorVoting with Jaccard distance Elizabeth Warren (MA) Dianne Feinstein (CA) Bernie Sanders (VT) d Mitch McConnell (KY) John McCain (AZ) Rand Paul (KY) Marco Rubio (FL) Ted Cruz (TX)
  • 18.
    Do votes alignwith bill sponsors? Republican Sponsor • The bill sponsor is the member of congress that introduces the document for consideration. Democrat Sponsor
  • 19.
    Infer directionality ofbiochemical reactions using Langevin dynamics Robert Vogel Developed new parameterization of therapeutic drugs using insight from nonlinear dynamical systems
  • 20.
    Voting Distributions andthe Simulated Senate • Sample 5000 experimental senates using parameters from data • Data exhibit a more diverse distribution then simulation • Potential next step, use the Ising model to model pairwise interactions Republican SponsorDemocrat Sponsor
  • 21.
    The Jaccard Indexand Political Polarity
  • 22.
    Jaccard Index forMeasuring Polarity • Jaccard Index measures the number identical votes between Senator i and Senator j normalized to total votes • Polarity index is the average Jaccard index between Senator i and all Senators in party R. Jij = |vi vj| |vi [ vj| JiR = 1 NR X j2R Jij
  • 23.
    Distribution of polarityindex • If party politics were not a factor, these distributions would overlap
  • 24.
    Jaccard Distance forSenator Clustering • Jij 1 the more similar Senator i votes to Senator j. • Hierarchical clustering utilizes a dissimilarity measure. Standard solution 1 - Jij dJ (i, j) = 1 |vi vj| |vi [ vj| | {z } Jij
  • 25.
    Votes are strictlypartisan • Fraction of votes along party line, most votes are partisan
  • 26.
  • 27.
    Topic modeling oflegislative summaries Wordspaceperbill TopicSpace Bills Topics T S = S’ Y N0 Congress person Topic Probabilities P T’ = P’ New bill in topic space Probability of vote P Y N 0 Prediction Clustering
  • 28.
    Can we makepredictions of senator votes from legislative documents? Topic 1 Topic 2 Topic 3 Senator 1 Vote Document 1 Document 2 Document N Topic M
  • 29.
    Can we makepredictions of senator votes from legislative documents? Topic 1 Topic 2 Topic 3 Senator 1 Vote Document 1 Document 2 Document N Topic M VoteSenator 2
  • 30.
    Can we makepredictions of senator votes from legislative documents? Topic 1 Topic 2 Topic 3 Senator 1 Vote Document 1 Document 2 Document N Topic M Senator L Vote VoteSenator 2
  • 31.
    Legislative document reductionto topics • 2559 legislative summaries • Constructed 6403 word basis from text by: • removing stop words (e.g and, that, this, a) • removing non-english words • stemming (e.g. rested equal to rest) • TF-IDF • Cosine similarity to group topics
  • 32.
    Legislative document reductionto topics • 2559 legislative summaries • Constructed 6403 word basis from text by: • removing stop words (e.g and, that, this, a) • removing non-english words • stemming (e.g. rested equal to rest) • TF-IDF • Cosine similarity to group topics • Result: No structure in bill data, more data needed! Documents Documents
  • 33.
    Document dimensionality reductionnot sufficient with PCA 95% of the variablity corresponds to > 1000 dimensions A small topic space, represents a small portion of the variability
  • 34.
    tSNE dimensionality reductionsuggests no structure in bill data • Each point is a document in the reduced space defined by tSNE • t-distributed Stochastic Neighborhood Embedding maps points from a high to a low dimensional space by minimizing the Kullback-Leibler Divergence (minimize information loss).
  • 35.
  • 36.
    Why only chooseBills and Amendments? • In general, these documents can become law • Other votes are for approving nominations for office and resolutions. • Resolutions can be very diverse as shown below.
  • 37.
    Graduate Research: Anoverview • Langevin Dynamics to: • figure out direction in biochemical reactions, and • testing isolation of a network motif. • Bifurcation analysis to identify: • nodes in a network sensitive to therapeutic inhibition
  • 38.
    Biochemical Noise • Flowcytometry measures the relative quantity of <= 12 biochemical species per cell at a rate of 20,000 cells per second. • Fluorescent molecules are coupled to antibodies that specifically bind to a biochemical species. • Quantity of molecules is proportional to fluorescent signal −2 −1 0 1 2 −4 −3 −2 −1 0 1 2 3 4 PMA 1 PMA 2 PMA 3 Log2 Normalized pMEK Log2NormalizedppERK PMA 1 PMA 2 PMA 3 PMA 1 PMA 2 PMA 3 −2 −1 0 1 2 −4 −3 −2 −1 0 1 2 3 4 PMA : 1 Log2 pMEK Log2ppERK PMA : 1 −2 −1 0 1 2 −4 −3 −2 −1 0 1 2 3 4 PMA : 2 Log2 pMEK Log2ppERK PMA : 2 −2 −1 0 1 2 −4 −3 −2 −1 0 1 2 3 4 PMA : 3 Log2 pMEK Log2ppERK PMA : 3
  • 39.
    Fluctuations break symmetry ofaverage measurements Variance of Y > X Y X𝜉x 𝜉y O O • Fluctuations from source node propagates to target X Y𝜉y 𝜉x O O Variance of X > Y True Model False Model
  • 40.
    Fluctuations break symmetry ofaverage measurements Variance of Y > X 0.2 0.3 0.4 0.5 0.6 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Cov(pMEK, ppERK) Variance True Model pMEK ppERK 0.2 0.3 0.4 0.5 0.6 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Cov(pMEK, ppERK) Variance False Model pMEK ppERK 0.2 0.3 0.4 0.5 0.6 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Cov(pMEK, ppERK) Variance True Model pMEK ppERK 0.2 0.3 0.4 0.5 0.6 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Cov(pMEK, ppERK) Variance False Model pMEK ppERK Y X𝜉x 𝜉y O O • Fluctuations from source node propagates to target X Y𝜉y 𝜉x O O Variance of X > Y True Model False Model
  • 41.
    Nonlinear dynamics ofbiochemical inhibition
  • 42.
    Inhibition of biochemicalsignaling in cells, a new parameter 𝛼 L c SRC pMEK MEKi ppERK In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
  • 43.
    Inhibition of biochemicalsignaling in cells, a new parameter 𝛼 L c SRC pMEK MEKi ppERK L c SRC SRCi pMEK ppERK In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
  • 44.
    Inhibition of biochemicalsignaling in cells, a new parameter 𝛼 L c SRC pMEK MEKi ppERK L c SRC SRCi pMEK ppERK In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
  • 45.
    Inhibition of biochemicalsignaling in cells, a new parameter 𝛼 L c SRC pMEK MEKi ppERK L c SRC SRCi pMEK ppERK In preparation for publicationChemical SpeciesChemical Complex Enzymatic reaction Enzymatic Inhibition
  • 46.
    Nonlinear dynamics of biochemicalinhibition in cells Chemical Species Chemical Complex Enzymatic reaction Enzymatic Inhibition In preparation for publication
  • 47.
    Nonlinear dynamics of biochemicalinhibition in cells Chemical Species Chemical Complex Enzymatic reaction Enzymatic Inhibition L c SRC pMEK MEKi ppERK [MEKi] [MEKi] In preparation for publication
  • 48.
    Nonlinear dynamics of biochemicalinhibition in cells Chemical Species Chemical Complex Enzymatic reaction Enzymatic Inhibition L c SRC SRCi pMEK ppERK [SRCi] [SRCi] L c SRC pMEK MEKi ppERK [MEKi] [MEKi] In preparation for publication
  • 49.
  • 50.
    Single cell measurementsfind abnormalities in tumor patient profiles • Kullback-Leibler divergence measures the dissimilarity of the single cell distribution of biochemical signaling features between patient and healthy donor samples. Sjk = X i2HD DKL (Pj(xk)||Pi(xk)) = X i2HD Pj(xk) log ✓ Pj(xk) Pi(xk) ◆ • k = Biochemical species • j = patient id • i = Healthy donor