Shows crude similarities of voting histories between members of Congress using roll call vote matrices from Poole, McCarty and Lewis: http://www.voteview.com/dwnl.htm. Uses vectorized operations to make similarity calculations happen super fast.

compare.py

'''
compare.py

Quickly produces a pairwise similarity matrix of lawmakers' roll call votes, given
an input *.ord matrix file from Poole, McCarty and Lewis: http://www.voteview.com/dwnl.htm
'''

import numpy, string
from scipy.spatial.distance import cdist

########## HELPERS ##########

class LookerUpper(object):
    '''
    Helps look up pairwise similarity scores by member, given an input name.
    '''
    def __init__(self, names, matrix):
        self.names = names
        self.matrix = matrix

    def lookup(self, name):
        for i, j in enumerate(self.matrix[self.names.index(name)]):
            yield '%s: %s' % (self.names[i], j)

    def lookup_pair(self, name1, name2):
        name1_idx = self.names.index(name1)
        name2_idx = self.names.index(name2)
        return '%s -> %s: %s' % (name1, name2, self.matrix[name1_idx, name2_idx])

def check(status):
    '''
    Zero out any non-votes, for Jaccard.
    '''
    if int(status) in [7, 8, 9]:
        return 0
    return status

def cleansplit(str):
    '''
    Split fixed-width input file. Sigh.
    '''
    return map(string.strip, [str[:12], str[12:20], str[:20:25], str[25:36], str[36:]])

########## MAIN ##########

if __name__ == '__main__':
    with open('hou112kh.ord.txt', 'rU') as infile:
        rows = [cleansplit(a) for a in infile.readlines()] # Parse fixed-width .ord input file
        names = [r[3] for r in rows] # r[3] is the name code

    # Build a numpy matrix from all the vote data in the input file.
    data = numpy.matrix([[check(i) for i in list(r[4])] for r in rows])

    # Calculate vectorized pairwise similarity between all senators using Jaccard distance,
    # which is just a measure of what percentage of each lawmaker's votes lined up with another's.
    # Subtracting from 1.0 just turns the measure into a similarity score rather than a distance.
    similarities = 1.0 - cdist(data, data, 'jaccard')

    # Print similarities, given a lookup name
    looker_upper = LookerUpper(names, similarities)
    for i in looker_upper.lookup('BOEHNER'):
        print i

    # And here's a lookup of two names
    print looker_upper.lookup_pair('MCCAUL', 'CANTOR')