def ReadEnvironment(remove_center=False):
# This method was used to read and create feature matrices from a previously
# created environment in a first stage of the project. This method might be outdated.
print("Extracting word features without n-grams...", end="")
(X_train_w, Y_train, X_test_w, Y_test, vectorizer,
feature_names_w) = features.ExtractWordFeatures(
"speeches_110_dwnominate_nonames.txt",
"speeches_112_dwnominate_nonames.txt",
remove_center_interval=remove_center)
print("[DONE]")
print("Extracting word features with 2-grams...", end="")
(X_train_w2, Y_train, X_test_w2, Y_test, vectorizer2,
feature_names_w2) = features.ExtractWordFeatures(
"speeches_110_dwnominate_nonames.txt",
"speeches_112_dwnominate_nonames.txt",
ngrams=2,
remove_center_interval=remove_center)
print("[DONE]")
print("Reading collocation features...", end="")
(X_train_c, feature_names_c) = features.ReadX("X_train_1000_l_2.txt")
(X_test_c, feature_names_c) = features.ReadX("X_test_1000_l_2.txt")
print("[DONE]")
if remove_center:
print("Removing center...", end="")
path_train = "../datasets/train/"
train_dataset_df = pd.read_csv(path_train +
"speeches_110_dwnominate_nonames.txt",
sep="|",
encoding="latin_1",
header=None)
train_dataset_df.columns = ['nominate_dim1', 'nominate_dim2', 'speech']
dw_nominates = train_dataset_df['nominate_dim1'].values.tolist()
indices = []
for i in range(0, len(dw_nominates)):
if float(dw_nominates[i]) < -0.2 or float(dw_nominates[i]) > 0.2:
indices.append(i)
X_train_c = [X_train_c[index] for index in indices]
print("[DONE]")
print("Joining word features without n-grams and collocations...", end="")
feature_names_t = feature_names_w + feature_names_c
X_train_t = hstack((X_train_w, X_train_c))
X_test_t = hstack((X_test_w, X_test_c))
print("[DONE]")
return X_train_w, Y_train, X_test_w, Y_test, vectorizer, feature_names_w, X_train_w2, X_test_w2, vectorizer2, feature_names_w2, X_train_c, X_test_c, feature_names_c, X_train_t, X_test_t, feature_names_t
def RemoveCenterExperimentBoW():
# This experiment evaluates the performance of the model in
# relation with the amount of moderate tests we remove. To do
# so, we calculate the accuracy of the model when we eliminate
# some percentage of the DW Nominate space around zero.
# Considering the highest positive and negative observations,
# we extract a 10% more of the Nomi-nate scale from zero to
# these values on each iteration, and calculate the accuracy
# over the rest of the observations.
centers = [
None, [-0.068, 0.091], [-0.136, 0.182], [-0.204, 0.273],
[-0.272, 0.364], [-0.34, 0.455], [-0.408, 0.546], [-0.476, 0.637],
[-0.544, 0.728], [-0.612, 0.819]
]
#centers = [[-0.2, 0.2]]
nb_accuracies = []
lr_accuracies = []
t0 = time.time()
t1 = time.time()
for i in range(0, len(centers)):
print(i)
X_train, Y_train, X_test, Y_test, vectorizer, feature_names = features.ExtractWordFeatures(
"speeches_110_dwnominate_nonames.txt",
"speeches_112_dwnominate_nonames.txt",
vectorizer_type="CountVectorizer",
ngrams=None,
balance_dataset=False,
remove_center_interval=centers[i])
print(len(Y_train))
print(len(Y_test))
print("NB")
nb = models.NBtrain(X_train, Y_train)
(accuracy, precision, recall) = evaluation.Test(X_test, Y_test, nb)
nb_accuracies.append(accuracy)
print("LR")
lr = models.LRtrain(X_train, Y_train)
accuracy, precision, recall = evaluation.Test(X_test, Y_test, lr)
lr_accuracies.append(accuracy)
t2 = time.time()
print(t2 - t1, " seconds")
t1 = time.time()
t3 = time.time()
print("Total time: ", t3 - t0)
x = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
plt.plot(x, nb_accuracies, color="#2D00B7", label="Naive Bayes")
plt.title("Accuracy and center removal relation")
plt.xlabel("Percentage of DW Nominate removed")
plt.ylabel("Accuracy")
plt.plot(x, lr_accuracies, color="#64EEBA", label="Logistic Regression")
plt.legend()
return nb_accuracies, lr_accuracies
def SetEnvironment():
# This method was used to create the train and test datasets and
# extract some feature matrices from this datasets in a
# first stage of the project. This method might be outdated.
print("Creating train dataset...")
CreateDataset("110_SpeakerMap.txt",
"speeches_110.txt",
"HS110_members.csv",
True,
"train_dataset.csv",
includeNames=False)
print("Creating test dataset...")
CreateDataset("112_SpeakerMap.txt",
"speeches_112.txt",
"HS112_members.csv",
False,
"test_dataset.csv",
includeNames=False)
print("Extracting word features...")
(X_train_w, Y_train_w, X_test_w, Y_test_w, vectorizer,
BoW_names) = features.ExtractWordFeatures("train_dataset.csv",
"test_dataset.csv")
print("Extracting collocation features...")
(X_train_1000, Y_train_1000, X_test_1000, Y_test_1000,
feature_names_1000) = features.ExtractCollocationFeatures(
"train_dataset.csv", "test_dataset.csv", "X_train_1000.txt",
"X_test_1000.txt", 5, 1000)
print("Creating total matrices...")
feature_words = vectorizer.get_feature_names()
total_features = feature_words + feature_names_1000
X_train_t = hstack((X_train_w, X_train_1000))
X_test_t = hstack((X_train_w, X_train_1000))
return X_train_w, Y_train_w, X_test_w, Y_test_w, X_train_1000, Y_train_1000, X_test_1000, Y_test_1000, X_train_t, X_test_t, vectorizer, feature_words, feature_names_1000, total_features
def crossValidationExperiment():
# This method implements a experiment in which we
# apply a 10-fold cross validation technique to the
# train dataset for both, Naive Bayes and Logistic Regression
# models.
(X_train, Y_train, X_test, Y_test, vectorizer,
feature_names) = features.ExtractWordFeatures(
"speeches_110_dwnominate_nonames.txt",
"speeches_112_dwnominate_nonames.txt",
vectorizer_type="CountVectorizer",
ngrams=None,
balance_dataset=False,
remove_center_interval=None)
nb = MultinomialNB()
scores_nb = cross_val_score(nb, X_train, Y_train, cv=10)
score_nb = np.average(scores_nb)
print("Accuracy NB: ", score_nb)
lr = LogisticRegression(solver='lbfgs', max_iter=1000)
scores_lr = cross_val_score(lr, X_train, Y_train, cv=10)
score_lr = np.average(scores_lr)
print("Accuracy LR: ", score_lr)
def differentSetsCominationsExperiment():
# In this experiment we introduce more congresses among
# our datasets. We try different combinations as
# train-test datasets including data from 109, 110, 111 and 112 Congresses.
# We carry out the experiment always with the the BoW model and
# with Naïve Bayes and Logistic Regression algorithms
congresses = ["109", "110", "111", "112"]
print("Only one congress for training.")
for i in range(0, len(congresses)):
train_congress = congresses[i]
for j in range(i + 1, len(congresses)):
test_congress = congresses[j]
print(train_congress, test_congress)
(X_train, Y_train, X_test, Y_test, vectorizer,
feature_names) = features.ExtractWordFeatures(
"speeches_" + train_congress + "_dwnominate_nonames.txt",
"speeches_" + test_congress + "_dwnominate_nonames.txt",
vectorizer_type="CountVectorizer",
ngrams=None,
balance_dataset=False,
remove_center_interval=None)
evaluation.Evaluate(
X_train, Y_train, X_test, Y_test,
"speeches_" + test_congress + "_dwnominate_nonames.txt",
feature_names)
print("Two congresses for training")
for i in range(0, len(congresses) - 1):
train_congress_1 = congresses[i]
train_congress_2 = congresses[i + 1]
train_dataset_df_1 = features.datasetTodf(
"../datasets/train/speeches_" + train_congress_1 +
"_dwnominate_nonames.txt")
train_dataset_df_2 = features.datasetTodf(
"../datasets/train/speeches_" + train_congress_2 +
"_dwnominate_nonames.txt")
train_dataset_df = pd.concat([train_dataset_df_1, train_dataset_df_2],
ignore_index=True)
for j in range(i + 2, len(congresses)):
test_congress = congresses[j]
test_dataset_df = features.datasetTodf(
"../datasets/test/speeches_" + test_congress +
"_dwnominate_nonames.txt")
print("Train:", train_congress_1, train_congress_2, "Test:",
test_congress)
(X_train, Y_train, X_test, Y_test, vectorizer,
feature_names) = features.ExtractWordFeaturesWithDataframes(
train_dataset_df,
test_dataset_df,
vectorizer_type="CountVectorizer",
ngrams=None,
balance_dataset=False,
remove_center_interval=None)
evaluation.Evaluate(
X_train, Y_train, X_test, Y_test,
"speeches_" + test_congress + "_dwnominate_nonames.txt",
feature_names)
print("Three congresses for training")
train_congresses = congresses[:3]
train_datasets = []
for congress in train_congresses:
train_dataset_df_i = features.datasetTodf(
"../datasets/train/speeches_" + congress +
"_dwnominate_nonames.txt")
train_datasets.append(train_dataset_df_i)
train_dataset_df = pd.concat(train_datasets, ignore_index=True)
test_congress = congresses[len(congresses) - 1]
test_dataset_df = features.datasetTodf("../datasets/test/speeches_" +
test_congress +
"_dwnominate_nonames.txt")
print("Train:", train_congresses, "Test:", test_congress)
(X_train, Y_train, X_test, Y_test, vectorizer,
feature_names) = features.ExtractWordFeaturesWithDataframes(
train_dataset_df,
test_dataset_df,
vectorizer_type="CountVectorizer",
ngrams=None,
balance_dataset=False,
remove_center_interval=None)
evaluation.Evaluate(
X_train, Y_train, X_test, Y_test,
"speeches_" + test_congress + "_dwnominate_nonames.txt", feature_names)