526 bill votes: 4 sessions of Congress; 174 commi:ees; 1156 keywords; 1414 related bills; 227 bill sponsors … Manually sorted keywords into the five categories displayed.
526 bill votes: 4 sessions of Congress; 174 commi:ees; 1156 keywords; 1414 related bills; 227 bill sponsors … Manually sorted keywords into the five categories displayed. Skewed dataset: Most bills that make it to a vote pass.
vote pass. The DATA Wide not deep: Of 526 bill votes: 4 sessions of Congress; 174 commi:ees; 1156 keywords; 1414 related bills; 227 bill sponsors … Manually sorted keywords into the five categories displayed.
but I model as a regression to get more insight • 10-‐fold cross valida5on • Precision is far more cri5cal than recall: Precision: 71%, Recall: 44% • Trim to top ~200 parameters for final model • Run that on bills that have been acDve in the last 3 weeks
live have both been correctly predicted: Passed at 58% (close) Intelligence Authoriza5on Act June 16, 2015 Passed at 67% (not close) Protect Medical Innova5on Act June 18, 2015
live have both been correctly predicted: Passed at 58% (close) Intelligence Authoriza5on Act June 16, 2015 Passed at 67% (not close) Protect Medical Innova5on Act June 18, 2015
live have both been correctly predicted: Passed at 58% (close) Intelligence Authoriza5on Act June 16, 2015 q House vote q Introduced by Rep. Devin Nunes (R-‐CA, male)
Regression: Tended more toward recall than precision Training RF on each topic’s bills only, vs. training on en5re dataset and tes5ng on 1/4th (via 4-‐fold CV) of topic bills. La:er performs be:er, so I stayed with global RF. I also did a grid search SVM with a β-‐ weighted F-‐score. β = .2 to encourage precision over recall. This did similarly well to RF, but without the straighmorward feature trimming.
for final model • Run that on bills that have been acDve in the last 3 weeks • I ul5mately want a classifier, but I model as a regression to get more insight. • 10-‐fold cross valida5on • Precision is far more cri5cal than recall.
and recall. • Looking for model improvement as parameter sets are incorporated. Trim final parameter set to ~200 parameters that are important over all 10 valida5on rounds – only these will be used to judge upcoming bills. Data processing & Algorithm: Random Forest