MachineLearning_Finegrained_Dataset.py

'''
(Messy Code)
Sentiment Analysis: Text Classification using Machine Learning algorithms
'''

import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocator
import itertools
import numpy as np
import pandas as pd
import string
import copy
from re import sub
from collections import defaultdict
from time import time

from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer, TfidfTransformer
from sklearn.model_selection import train_test_split, RepeatedStratifiedKFold, GridSearchCV
from sklearn.feature_extraction.stop_words import ENGLISH_STOP_WORDS
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.feature_selection import SelectKBest, chi2, SelectFromModel
from sklearn.decomposition import TruncatedSVD
from sklearn.naive_bayes import ComplementNB
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression, SGDClassifier
from sklearn.svm import LinearSVC
from sklearn.neural_network import MLPClassifier

from sklearn import metrics
from sklearn.datasets import load_files
from sklearn.externals import joblib

from nltk.corpus import stopwords
from nltk.corpus import wordnet as wn
from nltk.corpus import sentiwordnet as swn
from nltk import word_tokenize, sent_tokenize, pos_tag
from nltk.stem import WordNetLemmatizer


cross_validation_best = [0.0, 0.0, "", [], [], 0.0]           # [Accuracy, F1-score, Model Name, Actual Labels, Predicted, Time]
cross_validation_all = defaultdict(list)                      # {Name: (Accuracy, F1-score), (Accuracy, F1-score), ...}
cross_validation_average = defaultdict(list)                  # {Name: (Avg(Accuracy), Avg(F1-score)), ...}
time_complexity_average = defaultdict(list)                   # {Name: [Avg(Train+Test_Time)]

random_state = 22


def Run_Preprocessing(dataset_name):
    '''    Dataset Dependant Preprocessing    '''

    # 1. Dataset dependent loading
    data = ["" for i in range(294)]
    neg_count_feature = [0 for i in range(294)]  # 3 Extra columns, 3 Extra Features containing the counts (no sequential information)
    neu_count_feature = [0 for i in range(294)]
    pos_count_feature = [0 for i in range(294)]
    labels = ["" for i in range(294)]
    count = 0
    with open('./Datasets/Finegrained/finegrained.txt', 'r') as file:
        for line in file:
            if len(line.split("_")) == 3:
                labels[count] = line.split("_")[1]
            elif len(line.strip()) == 0:
                count += 1
            else:
                temp = [x.strip() for x in line.split("\t")]
                if len(temp[1]) > 1:
                    # "nr" label is ignored
                    if temp[0] == "neg":
                        neg_count_feature[count] += 1
                    elif temp[0] == "neu":
                        neu_count_feature[count] += 1
                    elif temp[0] == "pos":
                        pos_count_feature[count] += 1
                    elif temp[0] == "mix":
                        neg_count_feature[count] += 1
                        pos_count_feature[count] += 1        

                    data[count] += temp[1]

    print("--Processed", count+1, "documents", "\n--Dataset Name:", dataset_name)

    df = pd.DataFrame({'Labels': labels, 'Data': data, 'Neg_Feature': neg_count_feature, 'Neu_Feature': neu_count_feature, 'Pos_Feature': pos_count_feature})

    # 2. Remove empty instances from DataFrame, actually affects accuracy
    emptySequences = df.loc[df.loc[:,'Data'].map(len) < 1].index.values
    df = df.drop(emptySequences, axis=0).reset_index(drop=True)  # reset_Index to make the row numbers be consecutive again

    # 3. Shuffle the Dataset, just to make sure it's not too perfectly ordered
    if False:
        df = df.sample(frac=1, random_state=random_state).reset_index(drop=True)

    # 4. Print dataset information
    print("--Dataset Info:\n", df.describe(include="all"), "\n\n", df.head(3), "\n\n", df.loc[:,'Labels'].value_counts(), "\n--\n", sep="")

    # 5. Balance the Dataset by Undersampling
    if False:
        set_label = "neu"
        set_desired = 75

        mask = df.loc[:,'Labels'] == set_label
        df_todo = df[mask]
        df_todo = df_todo.sample(n=set_desired, random_state=random_state)
        df = pd.concat([df[~mask], df_todo], ignore_index=True)
        df = df.sample(frac=1, random_state=random_state).reset_index(drop=True)

    return df


# def Run_Classifier(grid_search_enable, pickle_enable, silent_enable, pipeline, parameters, data_train, data_test, labels_train, labels_test, targetnames, stopwords_complete_lemmatized, model_name):
#     '''    Run Classifier with or without Grid Search after Preprocessing is done    '''
#     time_counter = time()

#     ## GRID SEARCH ON - Search for the Best Parameters
#     if grid_search_enable == 1:

#         # (1) TRAIN
#         grid_go = GridSearchCV(pipeline, parameters, cv=3, n_jobs=-1)  # Run GridSearch in all Threads (Parallel)
#         grid_go = grid_go.fit(data_train, labels_train)
#         print('\n- - - - - BEST PARAMETERS - - - - -')
#         print(grid_go.best_score_, 'Accuracy')
#         for param_name in sorted(parameters.keys()):
#             print("%s: %r" % (param_name, grid_go.best_params_[param_name]))

#         print('\n- - - - - DETAILS - - - - -')
#         grid_results = grid_go.cv_results_['params']
#         for i in range(len(grid_results)):       
#             print(i, 'params - %s; mean - %0.10f; std - %0.10f' % (grid_results[i].values(), grid_go.cv_results_['mean_test_score'][i], grid_go.cv_results_['std_test_score'][i]))

#         # (2) Model Persistence (Pickle)
#         if pickle_enable == 1: joblib.dump(grid_go.best_estimator_, './pickled_models/Classifier.pkl')  

#         # (3) PREDICT
#         predicted = grid_go.predict(data_test)


#     ## GRID SEARCH OFF - Best Parameters are already known
#     else:   

#         # (1) TRAIN
#         pipeline.fit(data_train, labels_train)
#         if model_name not in ['(k-Nearest Neighbors)', '(Decision Tree)', '(Random Forest)', '(Multi-layer Perceptron)']: print('\nNumber of Features/Dimension is:', pipeline.named_steps['clf'].coef_.shape[1])
#         if model_name in ['(Decision Tree)']: print('\nNumber of Features/Dimension is:', pipeline.named_steps['clf'].n_features_, '| Tree Depth is:', pipeline.named_steps['clf'].tree_.max_depth)
#         if model_name in ['(Random Forest)']: print('\nNumber of Features/Dimension is:', pipeline.named_steps['clf'].n_features_)
#         if model_name in ['(Multi-layer Perceptron)']: print('\nNumber of Features/Dimension is:', pipeline.named_steps['clf'].coefs_[0].shape[0]) 

#         # (2) Model Persistence (Pickle)
#         if pickle_enable == 1: joblib.dump(pipeline, './pickled_models/Classifier.pkl') 

#         # (3) PREDICT
#         predicted = pipeline.predict(data_test)

#     Print_Result_Metrics(silent_enable, labels_test, predicted, targetnames, time_counter, 0, model_name)  


# def Print_Result_Metrics(silent_enable, labels_test, predicted, targetnames, time_counter, not_new_model, model_name):
#     '''    Print Metrics after Training (Testing phase)    '''
#     global cross_validation_best, cross_validation_all, time_complexity_average

#     # Times
#     if not_new_model == 0:
#         time_final = time()-time_counter
#     else:
#         time_final = time_counter

#     # Metrics
#     accuracy = metrics.accuracy_score(labels_test, predicted)
#     other_metrics_to_print = metrics.classification_report(labels_test, predicted, target_names=targetnames, output_dict=False)
#     other_metrics_as_dict = metrics.classification_report(labels_test, predicted, target_names=targetnames, output_dict=True)
#     confusion_matrix = metrics.confusion_matrix(labels_test, predicted)

#     if silent_enable == 0:
#         print('\n- - - - - RESULT METRICS -', "%.2fsec" % time_final, model_name, '- - - - -')
#         print('Exact Accuracy: ', accuracy)
#         print(other_metrics_to_print)
#         print(confusion_matrix)
#         print()

#     # Save to Global Variables
#     if not_new_model == 0:  # Lack of this is a fatal Bug; If this flag is 1 we are storing the same model twice
#         weighted_f1 = other_metrics_as_dict['weighted avg']['f1-score']
#         cross_validation_all[model_name].append((accuracy, weighted_f1))  # Tuple
#         time_complexity_average[model_name].append(time_final)

#         if accuracy > cross_validation_best[0]:
#             cross_validation_best[0] = accuracy
#             cross_validation_best[1] = weighted_f1
#             cross_validation_best[2] = model_name
#             cross_validation_best[3] = labels_test
#             cross_validation_best[4] = predicted 
#             cross_validation_best[5] = time_final         


# def Print_Result_CrossVal_Best(k):
#     '''    Print Metrics only of the best result that occured    '''
#     global cross_validation_best

#     if cross_validation_best[0] > 0.0:
#         print("\n" + "- " * 37, end = "")
#         Print_Result_Metrics(0, cross_validation_best[3], cross_validation_best[4], None, cross_validation_best[5], 1, cross_validation_best[2] + " best of " + str(k+1) + " Cross Validations")


# def Plot_Results(k, dataset_name):
#     '''    Plot the Accuracy of all Classifiers in a Graph    '''
#     global cross_validation_all, cross_validation_average

#     print("Plotting AVERAGES of Cross Validation...")
#     for model in cross_validation_all:
#         avg = tuple(np.mean(cross_validation_all[model], axis=0))
#         cross_validation_average[model] = avg  # Save the average on a global variable
#     indices = np.arange(len(cross_validation_average))
#     scores_acc = []
#     scores_f1 = []
#     model_names = []
#     for model in cross_validation_average:
#         scores_acc.append(cross_validation_average[model][0]) 
#         scores_f1.append(cross_validation_average[model][1])
#         model_names.append(model[1:-1])  # Remove Parentheses

#     # Reverse the items to appear in correct order
#     scores_acc.reverse()
#     scores_f1.reverse()
#     model_names.reverse()

#     fig, ax1 = plt.subplots(figsize=(15, 8))
#     fig.subplots_adjust(left=0.18, top=0.92, bottom=0.08)
#     fig.canvas.set_window_title(dataset_name + " - Averages across " + str(k) + "-fold Cross Validation")

#     p1 = ax1.bar(indices + 0.35, scores_acc, align="center", width=0.35, label="Accuracy (%)", color="navy")    
#     p2 = ax1.bar(indices, scores_f1, align="center", width=0.35, label="Accuracy (%)", color="cornflowerblue")

#     ax1.set_title(dataset_name + " - Averages across " + str(k) + "-fold Cross Validation")
#     ax1.set_ylim([0, 1])
#     ax1.yaxis.set_major_locator(MaxNLocator(11))
#     ax1.yaxis.grid(True, linestyle='--', which="major", color="grey", alpha=.25)
#     ax1.set_ylabel("Performance")    
#     ax1.legend((p1[0], p2[0]), ("Accuracy", "F1-score"))

#     ax1.set_xticks(indices + 0.35 / 2)
#     ax1.set_xticklabels(model_names)

#     # Rotates labels and aligns them horizontally to left 
#     plt.setp(ax1.xaxis.get_majorticklabels(), rotation=-45, ha="left", rotation_mode="anchor")

#     # Automatically adjust subplot parameters so that the the subplot fits in to the figure area
#     fig.tight_layout()

#     plt.show()
#     print()


class LemmaTokenizer(object):
    '''    Override SciKit's default Tokenizer    '''
    def __init__(self):
        self.wnl = WordNetLemmatizer()
        # This punctuation remover has the best Speed Performance
        self.translator = str.maketrans('','', sub('\'', '', string.punctuation))
    def __call__(self, doc):
        # return [self.wnl.lemmatize(t.lower()) for t in word_tokenize(doc)]
        temp = []
        for t in word_tokenize(doc):
            x = t.translate(self.translator) 
            if x != '': temp.append(self.wnl.lemmatize(x.lower())) 

        return temp


### START

# # Stopwords
# stopwords_complete = set(stopwords.words('english')).union(set(ENGLISH_STOP_WORDS))
# wnl = WordNetLemmatizer()
# stopwords_complete_lemmatized = set([wnl.lemmatize(word) for word in stopwords_complete])
# #

#np.set_printoptions(precision=10)  # Numpy Precision when Printing

df_dataset = Run_Preprocessing("Finegrained Sentiment Dataset")
all_data = df_dataset.loc[:,'Data']
all_labels = df_dataset.loc[:,'Labels']
neg_feature = df_dataset.loc[:,'Neg_Feature']
neu_feature = df_dataset.loc[:,'Neu_Feature']
pos_feature = df_dataset.loc[:,'Pos_Feature']

# Split using Cross Validation
set_fold = 5
cross_validation_enable = True  # Enable/Disable Flag; if disabled runs the evaluation just once
k_fold = RepeatedStratifiedKFold(5, n_repeats=1, random_state=random_state)


# Dimensionality Reduction - 4 different ways to pick the best Features 
#   (1) ('feature_selection', SelectKBest(score_func=chi2, k=5000)),  # 0.852 accuracy                    
#   (2) ('feature_selection', TruncatedSVD(n_components=1000)),  # Has Many Issues
#   (3) ('feature_selection', SelectFromModel(estimator=LinearSVC(), threshold=-np.inf, max_features=5000)),  # 0.860 accuracy 
#   (4) ('feature_selection', SelectFromModel(estimator=LinearSVC(penalty='l1', dual=False), threshold=-np.inf, max_features=5000)),  # 0.824 accuracy | Technically L1 is better than L2

metric_results_f1 = []
metric_results_acc = []

### (1) LET'S BUILD : Complement Naive Bayes
for k, (train_indexes, test_indexes) in enumerate(k_fold.split(all_data, all_labels)):  # Split must be done before every classifier because generated object gets exhausted (destroyed)
    print("\n--Current Cross Validation Fold:", k+1)

    data_train = all_data.reindex(train_indexes, copy=True, axis=0)
    labels_train = all_labels.reindex(train_indexes, copy=True, axis=0)
    neg_feature_train = neg_feature.reindex(train_indexes, copy=True, axis=0)
    neu_feature_train = neu_feature.reindex(train_indexes, copy=True, axis=0)
    pos_feature_train = pos_feature.reindex(train_indexes, copy=True, axis=0)

    data_test = all_data.reindex(test_indexes, copy=True, axis=0)
    labels_test = all_labels.reindex(test_indexes, copy=True, axis=0)
    neg_feature_test = neg_feature.reindex(test_indexes, copy=True, axis=0)
    neu_feature_test = neu_feature.reindex(test_indexes, copy=True, axis=0)
    pos_feature_test = pos_feature.reindex(test_indexes, copy=True, axis=0)


    # Grid Search Off
    # pipeline = Pipeline([ # Optimal
    #                     ('union', FeatureUnion(transformer_list=[      
    #                         ('vect1', CountVectorizer(max_df=0.90, min_df=5, ngram_range=(1, 1), stop_words=stopwords_complete_lemmatized, strip_accents='unicode', tokenizer=LemmaTokenizer())),  # 1-Gram Vectorizer
    #                         ('vect2', CountVectorizer(max_df=0.95, min_df=8, ngram_range=(2, 2), stop_words=None, strip_accents='unicode', tokenizer=LemmaTokenizer())),],  # 2-Gram Vectorizer

    #                         transformer_weights={
    #                             'vect1': 1.0,
    #                             'vect2': 1.0,},
    #                     )),
    #                     ('tfidf', TfidfTransformer(use_idf=True)),
    #                     #('feature_selection', SelectKBest(score_func=chi2, k=5000)),  # Dimensionality Reduction                  
    #                     ('clf', ComplementNB()),])  

    #Run_Classifier(0, 0, 1, pipeline, {}, data_train, data_test, labels_train, labels_test, None, stopwords_complete_lemmatized, '(Complement Naive Bayes)')

    # Training    
    vectorizer = TfidfVectorizer(max_df=0.92, min_df=5, ngram_range=(1, 1), stop_words="english", use_idf=True, strip_accents='unicode')
    tfidf_matrix = vectorizer.fit_transform(data_train)
    tfidf_dense = tfidf_matrix.toarray()  # dense ndarray
    print(type(tfidf_dense), tfidf_dense.shape)
    print(tfidf_dense.nonzero())

    #b = np.array([[100, 200], [100, 200], [100, 200], [100, 200], [100, 200], [100, 200]])
    #b = b[:, None]
    #print("The shapes are:",tfidf_dense.shape, b.shape) 
    #success = np.column_stack((tfidf_dense, b))  # or hstack
    neg_feature_train = neg_feature_train.values[:, None]
    neu_feature_train = neu_feature_train.values[:, None]    
    pos_feature_train = pos_feature_train.values[:, None]  
    print(tfidf_dense.shape)
    mode = "none"
    if mode == "Features":
        tfidf_dense = np.hstack((neg_feature_train, neu_feature_train, pos_feature_train))  # or column_stack 
    elif mode == "BoW_plus_Features":
        tfidf_dense = np.hstack((tfidf_dense, neg_feature_train, neu_feature_train, pos_feature_train))  # or column_stack  

    print(tfidf_dense.shape)
    #quit()

    from sklearn.neighbors import KNeighborsClassifier
    from sklearn.ensemble import RandomForestClassifier

    #clf = KNeighborsClassifier(40)
    clf = RandomForestClassifier(max_depth=None)
    #clf = RandomForestClassifier(max_depth=None, max_features=220)
    #clf = ComplementNB() # 1
    #clf = DecisionTreeClassifier(random_state=random_state) # 2
    #clf = LogisticRegression(penalty='l2', solver='lbfgs', multi_class='multinomial', max_iter=1000, C=1.0, n_jobs=1, random_state=random_state) # 3, solver : str, {‘newton-cg’, ‘lbfgs’, ‘liblinear’, ‘sag’, ‘saga’}, default: ‘liblinear’.
    #clf = LinearSVC(penalty='l2', max_iter=1000, dual=True, random_state=random_state) # 4

    
    clf.fit(tfidf_dense, labels_train)

    #Tree stuff
    #print(clf.max_features_)
    #from sklearn import tree
    #tree.export_graphviz(clf, out_file='tree.dot')   
    #quit()


    # Test
    tf_idf_matrix_test = vectorizer.transform(data_test)
    tfidf_dense_test = tf_idf_matrix_test.toarray()  # dense ndarray

    neg_feature_test = neg_feature_test.values[:, None]
    neu_feature_test = neu_feature_test.values[:, None]    
    pos_feature_test = pos_feature_test.values[:, None]  
    print(tfidf_dense_test.shape)
    if mode == "BoW_plus_Features":
        tfidf_dense_test = np.hstack((tfidf_dense_test, neg_feature_test, neu_feature_test, pos_feature_test))  # or column_stack  
    elif mode == "Features":
        tfidf_dense_test = np.hstack((neg_feature_test, neu_feature_test, pos_feature_test))  # or column_stack 
    print(tfidf_dense_test.shape)

    print(tfidf_dense_test.nonzero())

    predicted = clf.predict(tfidf_dense_test)


    accuracy = metrics.accuracy_score(labels_test, predicted)
    other_metrics_to_print = metrics.classification_report(labels_test, predicted, output_dict=False)
    other_metrics_as_dict = metrics.classification_report(labels_test, predicted, output_dict=True)
    confusion_matrix = metrics.confusion_matrix(labels_test, predicted)

    print('\n- - - - - RESULT METRICS - - - - - -')
    print('Exact Accuracy: ', accuracy)
    print(other_metrics_to_print)
    print(confusion_matrix)
    print()

    metric_results_f1.append(other_metrics_as_dict['weighted avg']['f1-score'])
    metric_results_acc.append(accuracy)


    #Print_Result_Metrics(0, labels_test, predicted, None, 0.1, 0, "ComplementNB")  

    if cross_validation_enable == False:
        break  # Disable Cross Validation

#print(clf.feature_importances_)

#Print_Result_CrossVal_Best(k)
print("F1 avg", np.around(np.mean(metric_results_f1)*100.0, decimals=3))
print("Acc avg", np.around(np.mean(metric_results_acc)*100.0, decimals=3))
print("F1 max", np.around(max(metric_results_f1)*100.0, decimals=3))
print("Acc max", np.around(max(metric_results_acc)*100.0, decimals=3))

###
quit()




Print_Result_CrossVal_Best(k)
###

# Don't print, just plot the averages of all models
Plot_Results(set_fold, "Finegrained Sentiment Dataset")