def feature_selection_trials():
"""
Select top k features. Vary k and plot data
"""
global pos, neg, totals, features
retrain = True
if not retrain and os.path.isfile(FDATA_FILE):
pos, neg, totals = cPickle.load(open(FDATA_FILE))
return
words = list(set(pos.keys() + neg.keys()))
print "Total no of features:", len(words)
words.sort(key=lambda w: -MI(w))
num_features, accuracy = [], []
bestk = 0
limit = 500
path = "./aclImdb/test/"
step = 500
start = 20000
best_accuracy = 0.0
for w in words[:start]:
features.add(w)
for k in xrange(start, 40000, step):
for w in words[k:k+step]:
features.add(w)
correct = 0
size = 0
for file in os.listdir(path + "pos")[:limit]:
correct += classify(open(path + "pos/" + file).read()) == True
size += 1
for file in os.listdir(path + "neg")[:limit]:
correct += classify(open(path + "neg/" + file).read()) == False
size += 1
num_features.append(k+step)
accuracy.append(correct / size)
if (correct / size) > best_accuracy:
bestk = k
print k+step, correct / size
features = set(words[:bestk])
cPickle.dump(get_relevant_features(), open(FDATA_FILE, 'w'))
pylab.plot(num_features, accuracy)
pylab.show()
def feature_selection_experiment(test_set):
"""
Select top k features. Vary k from 1000 to 50000 and plot data
"""
keys = positive.keys() + negative.keys()
sorted_keys = sorted(keys, cmp=lambda x, y: mutual_info(x) > mutual_info(y)) # Sort descending by mutual info
features = set()
num_features, accuracy = [], []
print sorted_keys[-100:]
for k in xrange(0, 50000, 1000):
features |= set(sorted_keys[k:k+1000])
preprocessor = partial(reduce_features, features)
correct = 0
for text, label in test_set:
correct += classify(text) == label
num_features.append(k+1000)
accuracy.append(correct / len(test_set))
print negate_sequence("Is this a good idea")
print reduce_features(features, "Is this a good idea")
pylab.plot(num_features, accuracy)
pylab.show()
