learning.py

#
# python run.py train.txt test.txt
#
# The train.txt and test.txt must be of the following format:
# 0 1 0 0 ... 0 0 1 0 : 0

import sys
import glob
import os
import re
import time

import numpy as np

from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier
from sklearn.naive_bayes import GaussianNB
from sklearn.lda import LDA
from sklearn.qda import QDA
from sklearn.linear_model import SGDRegressor
from sklearn import linear_model

#
# Checking command line arguments
#

f_train = ""
f_test  = ""

if len(sys.argv) == 3:
  f_train = sys.argv[1]
  f_test = sys.argv[2]
else:
  print "Usage: python learning.py <train-data> <test-data>"
  quit()

#
# Training Phase
#

features  = []
solutions = []
traindata = []

num_dups = 0
num_data = 0
num_cflct = 0

# read without duplications
with open(f_train) as f:
  for line in f:
    data = re.split(r':',line)
    proved = float(data[1].strip())
    wv = data[0]
    wv = re.split(r' |\t|\n', wv)
    wv = filter(lambda x:x.strip() != '', wv)
    wv = map(lambda x:int(x),wv)
    if (wv,proved) not in traindata:
      traindata.append((wv,proved))
    else:
      num_dups = num_dups + 1
    num_data = num_data + 1

print "Data without duplicates:"
print ""

for i in range(len(traindata)):
  (f,s) = traindata[i]
  print s, f
  features.append(f)
  solutions.append(s)

print ""
print "#data       : %d" % num_data
print "#duplucates : %d" % num_dups
print "#final-training-data : %d" % (len(traindata))


#
# Read test data
#

testFeatures = []
testSolutions = []
with open(f_test) as f:
  for line in f:
    data = re.split(r':',line)
    proved = float(data[1].strip())
    wv = data[0]
    wv = re.split(r' |\t|\n', wv)
    wv = filter(lambda x:x.strip() != '', wv)
    wv = map(lambda x:int(x),wv)
    testFeatures.append (wv)
    testSolutions.append(proved)

#
# prediction
#

C = 1.0
classifiers = [('Nearest Neighbors', KNeighborsClassifier(5)),
               ('Logistic Regression (l1)', LogisticRegression(C=C, penalty='l1')),
               ('Logistic Regression (l2)', LogisticRegression(C=C, penalty='l2')),
               ('Linear SVM', SVC(kernel='linear',C=C,probability=True,random_state=0)),
               ('RBF SVM', SVC(gamma=2,C=C,probability=True,random_state=0)),
               ('Decision Tree', DecisionTreeClassifier(max_depth=10)),
               ('Random Forest', RandomForestClassifier(max_depth=10,n_estimators=10,max_features=1)),
#               ('AdaBoost', AdaBoostClassifier()),
               ('Naive Bayes', GaussianNB())]

index = 0;

for (name, clf) in classifiers:
  
  start_time = time.time()

  index = index + 1
  clf.fit (features,solutions)

  print ""
  print "%d" % index, 
  print ".", name
  print clf

  correct = 0
  incorrect = 0
  true_total = 0
  false_total = 0
  true_hits = 0
  false_hits = 0

  for i in range(len(testFeatures)):
    pred = clf.predict(testFeatures[i])[0]
    solution = testSolutions[i]
    if pred == solution:
      correct = correct + 1
    else:
      incorrect = incorrect + 1
    if solution == 1:
      true_total = true_total + 1
      if pred == 1:
        true_hits = true_hits + 1
    if solution == 0:
      false_total = false_total + 1
      if pred == 0:
        false_hits = false_hits + 1
  
  total = correct + incorrect
  
 
  
  print "True  hits : %d / %d (%d%%)" % (true_hits, true_total, true_hits * 100 / true_total)
  print "False hits : %d / %d (%d%%)" % (false_hits, false_total, false_hits * 100 / false_total)
  print "Total      : %d / %d (%d%%)" % (correct,total,correct * 100 / total)

  end_time = time.time()

  print "%.2f" % (end_time - start_time),
  print "sec"

#  if hasattr(clf, "feature_importances_"):
#    importances = clf.feature_importances_
#    print importances