nodeImpurity.py

###
# Attribute Branching Impurity Calculator
# Zachary Friss
# CSCI 420
# Written for Python 3.4
###

from statistics import *
from math import *

def main():

  # File to parse
  fileName = "Storm_Weather_Data_v112.csv"

  # Define our attributes and information about them.
  f = open(fileName, "r")
  data = f.read()
  rows = data.split('\n')
  column_headers = []
  for row in rows:
    split_row = row.split(",")
    for field in split_row:
      column_headers.append(field)
  index = 0
  for field in column_headers:
    dataAttributes[index] = {"string": field, "index": index, "nodes": {}}
    
  # Index of target value
  targetValueIndex = 5

  # Target Value "Yes/True/Positive" value
  targetYesValue = "yes"

  ###
  # Read the file line by line. Remove any extra spaces or new lines.
  # Skip first line that has headings.
  # Split up the values by the comma.
  # Parse everything as an int and place it into data array.
  ###
  data = []
  firstLine = True
  for line in open(fileName):
    if firstLine:
      firstLine = False
      continue
    line = line.strip().split(',')
    entry = []
    for attribute in line:
      entry.append(attribute)
    data.append(entry)

  # Define totalEntries in data for weights
  totalEntries = len(data)

  # Iterate through the various attributes defined.
  # Calculate splits based on the attribute
  for attribute in dataAttributes:
    index = attribute['index']
    nodes = attribute['nodes']

    for entry in data:
      value = str(entry[index])
      targetYes = entry[targetValueIndex]

      # Value is something new initialize it.
      if not value in nodes:
        nodes[value] = {"targetYes": 0, "targetNo": 0}
        nodes[value]['value'] = value

      # Increment correct end value for datay entry.
      if targetYes == targetYesValue:
        nodes[value]["targetYes"] += 1
      else:
        nodes[value]["targetNo"] += 1

    mixedGiniIndex = 0
    mixedMisClassificationError = 0
    mixedEntropy = 0
    nodesArray = []

    # Iterate through nodes for attribute
    # Calculate the individual node's value
    for nodeValue, node in nodes.items():
      total = node['targetYes'] + node['targetNo']
      weight = total / totalEntries

      probTargetYes = node['targetYes'] / total
      probTargetNo = node['targetNo'] / total

      # Calculate Miscallsification Error
      misClassificationError = 1 - max([probTargetYes, probTargetNo])
      weightedMisClassificationError = weight * misClassificationError
      mixedMisClassificationError += weightedMisClassificationError

      # Calculate GINI
      probTargetYesSquared = pow(probTargetYes, 2)
      probTargetNoSquared = pow(probTargetNo, 2)

      gini = 1 - (probTargetYesSquared + probTargetNoSquared)
      weightedGini = weight * gini
      mixedGiniIndex += weightedGini

      # Calculate Entropy
      entropy = (-1 * (probTargetYes) * log2(probTargetYes)) + (-1 * (probTargetNo) * log2(probTargetNo))
      weightedEntropy = weight * entropy
      mixedEntropy += weightedEntropy
      node['entropy'] = entropy

      nodesArray.append(node)

    # Sort based on entropy
    nodesArray.sort(key=lambda n: n['entropy'], reverse = True)

    # Print out the branch's values
    print("Branching on " + str(attribute['string']))
    print("Weighted GINI index: "+str(round(mixedGiniIndex, 3)))
    print("Weighted Entropy: "+str(round(mixedEntropy, 3)))
    print("Weighted Misclassification error: "+str(round(mixedMisClassificationError, 3)))

    print("Individual Node Values:")
    lastIndex = len(nodesArray) - 1
    for index, node in enumerate(nodesArray):
      if index == 0:
        print("Highest Entropy: (" + str(node['value'])+") " + str(node['entropy']))
      elif index == lastIndex:
        print("Lowest Entropy: (" + str(node['value'])+") "  + str(node['entropy']))
      else:
        print("Middle Entropy: (" + str(node['value'])+") "  + str(node['entropy']))


    print()


main()