All files

This includes all files used for model development as well as the final submission files. Note that the raw data is not included.

README.md

@@ -1,2 +1,6 @@
 # Amazon product rating
  Project for rating products on amazon based on reviewer/product data
+
+Please see https://kasra-eshaghi.github.io/blog/software/amazon-product-rating for more details.
+
+The data for this project can be found at https://www.kaggle.com/c/csc2515-rating-prediction 

all_models.py

@@ -0,0 +1,183 @@
+# -*- coding: utf-8 -*-
+
+import tensorflow as tf
+from tensorflow import keras
+
+def model_1(train_review_time, train_categories, train_summaries, train_reviews, cv_summary_fit, cv_review_fit):
+    # model 1 - min val = 0.5619
+    input_time = keras.layers.Input(shape = [train_review_time.shape[1]], name='time_input')
+    input_cat = keras.layers.Input(shape = [train_categories.shape[1]], name='cat_input')
+    input_summary = keras.layers.Input(shape = [train_summaries.shape[1]], name='summary_input')
+    input_review = keras.layers.Input(shape = [train_reviews.shape[1]], name='review_input')
+
+    embed_summary = keras.layers.Embedding(input_dim = (len(cv_summary_fit.vocabulary_)+1), output_dim = 5, input_length = train_summaries.shape[1], name = 'embed_summary')(input_summary)
+    embed_review = keras.layers.Embedding(input_dim = (len(cv_review_fit.vocabulary_)+1), output_dim = 32, input_length = train_reviews.shape[1], name = 'embed_review')(input_review)
+
+    flatten_summary = keras.layers.Flatten()(embed_summary)
+    flatten_review = keras.layers.Flatten()(embed_review)
+
+    concat_all = keras.layers.concatenate([input_time, input_cat, flatten_summary, flatten_review])
+
+    hidden_1 = keras.layers.Dense(units = 100, activation = 'relu', name = 'hidden')(concat_all)
+
+    output = keras.layers.Dense(units = 1, name = 'output')(hidden_1)
+
+    model = keras.Model(inputs=[input_time, input_cat, input_summary, input_review], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+    return model
+
+def model_2(train_reviews, cv_review_fit):
+    #min val = 0.6
+    model = keras.models.Sequential([
+    keras.layers.Embedding(input_dim = (len(cv_review_fit.vocabulary_)+1), output_dim = 32, input_length = train_reviews.shape[1], name = 'embed_review'),
+    keras.layers.LSTM(units = 128),
+    keras.layers.Dense(units = 1, name = 'output'),
+    ])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+    return model
+
+def model_3(train_review_time, train_categories, train_summaries, train_reviews, cv_summary_fit, cv_review_fit):
+    # model 3 - min val = 0.58
+    input_time = keras.layers.Input(shape = [train_review_time.shape[1]], name='time_input')
+    input_cat = keras.layers.Input(shape = [train_categories.shape[1]], name='cat_input')
+    input_summary = keras.layers.Input(shape = [train_summaries.shape[1]], name='summary_input')
+    input_review = keras.layers.Input(shape = [train_reviews.shape[1]], name='review_input')
+
+    embed_cat = keras.layers.Embedding(input_dim = 5, output_dim = 5, input_length = train_categories.shape[1], name = 'embed_cat')(input_cat)
+    embed_summary = keras.layers.Embedding(input_dim = (len(cv_summary_fit.vocabulary_)+1), output_dim = 5, input_length = train_summaries.shape[1], name = 'embed_summary')(input_summary)
+    embed_review = keras.layers.Embedding(input_dim = (len(cv_review_fit.vocabulary_)+1), output_dim = 32, input_length = train_reviews.shape[1], name = 'embed_review')(input_review)
+
+    flatten_cat = keras.layers.Flatten()(embed_cat)
+    flatten_summary = keras.layers.Flatten()(embed_summary)
+    flatten_review = keras.layers.Flatten()(embed_review)
+
+    concat_all = keras.layers.concatenate([input_time, flatten_cat, flatten_summary, flatten_review])
+
+    hidden_1 = keras.layers.Dense(units = 100, activation = 'relu', name = 'hidden')(concat_all)
+
+    output = keras.layers.Dense(units = 1, name = 'output')(hidden_1)
+
+    model = keras.Model(inputs=[input_time, input_cat, input_summary, input_review], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+    return model
+
+def model_4(train_review_time, train_categories, train_summaries, train_reviews, cv_summary_fit, cv_review_fit):
+    # model 4 - min val = 0.86
+    input_time = keras.layers.Input(shape = (train_review_time.shape[1],), name='time_input')
+    input_cat = keras.layers.Input(shape = (train_categories.shape[1],), name='cat_input')
+
+    embed_cat = keras.layers.Embedding(input_dim = 5, output_dim = 5, input_length = train_categories.shape[1], name = 'embed_cat')(input_cat)
+    flatten_cat = keras.layers.Flatten(name = 'flatten_cat')(embed_cat)
+    concat_time_cat = keras.layers.concatenate([input_time, flatten_cat], name = 'concat_time_cat')
+    dense_time_cat = keras.layers.Dense(units = 10, activation = 'relu')(concat_time_cat)
+
+
+    input_summary = keras.layers.Input(shape = (train_summaries.shape[1],), name='summary_input')
+    embed_summary = keras.layers.Embedding(input_dim = (len(cv_summary_fit.vocabulary_)+1), output_dim = 32, input_length = train_summaries.shape[1], name = 'embed_summary')(input_summary)
+    lstm_summary = keras.layers.LSTM(units = 128)(embed_summary)
+    dense_summary = keras.layers.Dense(units = 1)(lstm_summary)
+
+    input_review = keras.layers.Input(shape = (train_reviews.shape[1],), name='review_input')
+    embed_review = keras.layers.Embedding(input_dim = (len(cv_review_fit.vocabulary_)+1), output_dim = 32, input_length = train_reviews.shape[1], name = 'embed_review')(input_review)
+    lstm_review = keras.layers.LSTM(units = 128)(embed_review) 
+    dense_review = keras.layers.Dense(units = 1)(lstm_review)   
+
+    concat_all = keras.layers.concatenate([dense_time_cat, dense_summary, dense_review])
+
+
+    output = keras.layers.Dense(units = 1, name = 'output')(concat_all)
+
+    model = keras.Model(inputs=[input_time, input_cat, input_summary, input_review], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+    return model
+
+def model_5(train_review_time, train_categories, train_summaries, train_reviews, cv_summary_fit, cv_review_fit):
+    # model 5 - min val =0.55
+    input_time = keras.layers.Input(shape = [train_review_time.shape[1]], name='time_input')
+    input_cat = keras.layers.Input(shape = [train_categories.shape[1]], name='cat_input')
+    input_summary = keras.layers.Input(shape = [train_summaries.shape[1]], name='summary_input')
+    input_review = keras.layers.Input(shape = [train_reviews.shape[1]], name='review_input')
+
+    embed_summary = keras.layers.Embedding(input_dim = (len(cv_summary_fit.vocabulary_)+1), output_dim = 16, input_length = train_summaries.shape[1], name = 'embed_summary')(input_summary)
+    embed_review = keras.layers.Embedding(input_dim = (len(cv_review_fit.vocabulary_)+1), output_dim = 32, input_length = train_reviews.shape[1], name = 'embed_review')(input_review)
+
+    conv1d_summary = keras.layers.Conv1D(filters = 128, kernel_size=5)(embed_summary)
+    conv1d_review = keras.layers.Conv1D(filters = 128, kernel_size=5)(embed_review)
+
+    flatten_summary = keras.layers.Flatten()(conv1d_summary)
+    flatten_review = keras.layers.Flatten()(conv1d_review)
+
+    concat_all = keras.layers.concatenate([input_time, input_cat, flatten_summary, flatten_review])
+
+    hidden_1 = keras.layers.Dense(units = 100, activation = 'relu', name = 'hidden')(concat_all)
+
+    output = keras.layers.Dense(units = 1, name = 'output')(hidden_1)
+
+    model = keras.Model(inputs=[input_time, input_cat, input_summary, input_review], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+    return model
+
+def model_6(train_summaries, train_reviews, vocab_size, max_sent_len):
+    # create model, val error = 0.63
+    input_reviews = keras.layers.Input(shape = (train_reviews.shape[1],), name = 'review_input')
+    input_summaries = keras.layers.Input(shape = (train_summaries.shape[1],), name = 'summary_input')
+
+    embed_words = keras.layers.Embedding(input_dim = vocab_size, output_dim = 8, input_length = max_sent_len, mask_zero=True, name = 'word_embedding')
+
+    input_reviews_encoded = embed_words(input_reviews)
+    input_summaries_encoded = embed_words(input_summaries)
+
+
+    concat_embeded_sentences = keras.layers.Concatenate(axis = -1, name = 'summary_and_review')([input_reviews_encoded, input_summaries_encoded])
+
+    lstm_sentences = keras.layers.LSTM(64, return_sequences = False, name = 'lstm_out')(concat_embeded_sentences)
+
+    output = keras.layers.Dense(1)(lstm_sentences)
+
+    model = keras.Model(inputs=[input_reviews, input_summaries], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+def model_7(train_reviews, vocab_size, max_sent_len):
+    #%% create model, val error = 0.5
+    input_reviews = keras.layers.Input(shape = (train_reviews.shape[1],), name = 'review_input')
+
+    embed_review = keras.layers.Embedding(input_dim = vocab_size, output_dim = 32, input_length = max_sent_len, mask_zero=False)(input_reviews)
+
+    gru_review = keras.layers.GRU(units = 128, return_sequences=False)(embed_review)
+    #gru_review_2 = keras.layers.GRU(units = 128)(gru_review)
+
+    output = keras.layers.Dense(1)(gru_review)
+
+    model = keras.Model(inputs=[input_reviews], outputs = [output])
+
+    model.compile(loss = 'mse', optimizer = 'Adam') 
+
+def build_model(n_hidden = 1, n_neurons = 50, learning_rate = 3e-3,  inputshape= 5):
+    model = keras.models.Sequential()
+    # add input layer
+    model.add(keras.layers.InputLayer(input_shape = (inputshape,)))
+
+    model.add(keras.layers.Dense(n_neurons, activation = 'relu', input_shape = (inputshape,)))
+    # add desired number of hidden layers:
+    for layer in range(n_hidden-1):
+        model.add(keras.layers.Dense(n_neurons, activation = 'relu'))
+
+    # add last dense layer to get output:
+    model.add(keras.layers.Dense(1))
+
+    # compile model:
+    model.compile(loss ='mse', optimizer = keras.optimizers.SGD(lr = learning_rate))
+
+    return model
+

data_analysis.py

@@ -0,0 +1,147 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Dec  6 17:43:10 2020
+
+@author: Kasra
+"""
+
+import pandas as pd
+import matplotlib.pyplot as plt
+import numpy as np
+
+from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer
+from nltk.corpus import stopwords
+from nltk.stem import WordNetLemmatizer, PorterStemmer 
+
+# import data
+df = pd.read_json('train.json', lines=True)
+# drop unnecessary columns
+df.drop('image', inplace = True, axis = 1)
+df.drop('reviewTime', inplace = True, axis = 1)
+df.drop('reviewHash', inplace = True, axis = 1)
+
+# clean up price data:
+df['price'] = df['price'].replace({'\$':''}, regex = True)
+
+# clean up review text and summary text:
+df['reviewText'] = df['reviewText'].replace(np.nan,'', regex = True)
+df['summary'] = df['summary'].replace(np.nan,'', regex = True)
+
+#%% Look at rating for each music category over time
+categories = df['category'].unique()
+for category in categories:
+    colors = ['red', 'blue', 'green', 'magenta','gold']
+    data = []
+    for overall in range(1,6,1):
+        # find dataframe for specific overall and category:
+        temp_df = df.loc[(df['category'] == category) & (df['overall'] == overall)]
+        data.append(temp_df.unixReviewTime)
+
+
+
+    plt.hist(data, color = colors, label = ['1','2','3','4','5'], bins = [0.9e9, 1e9, 1.1e9, 1.2e9, 1.3e9, 1.4e9, 1.5e9, 1.6e9])
+    plt.title(category)
+    plt.xlim(0.9e9, 1.6e9)
+    plt.ylim(0, 30000)
+    plt.xlabel('unixReviewTime ')
+    plt.ylabel('Frequency')
+
+    plt.show()
+
+#%% Look at statistics for review and summary data:
+cv = CountVectorizer(ngram_range = (1,1), lowercase = True, analyzer = 'word', binary = True)
+analyzer = cv.build_analyzer()
+
+# count number of words in the summary and reviews
+summary_texts = list(df['summary'])
+summary_lengths = []
+for summary in summary_texts:
+    # tokenize:
+    summary_tok = analyzer(summary)
+    summary_lengths.append(len(summary_tok))
+
+review_texts = list(df['reviewText'])
+review_lengths = []
+for review in review_texts:
+    # tokenize:
+    review_tok = analyzer(review)
+    review_lengths.append(len(review_tok))
+
+plt.boxplot([summary_lengths], vert = False, labels = ['Summary'])
+plt.title('Summary Length')
+plt.xlim(0,40)
+plt.show()
+
+plt.boxplot([review_lengths], vert = False, labels = ['Review'])
+plt.title('Review Length')
+plt.xlim(0, 5500)
+plt.show()
+
+# count number of unique words in summary and review
+cv_fit_summary = cv.fit(summary_texts)
+print('number of unique words in summaries:', len(cv_fit_summary.vocabulary_.keys()))
+
+cv_fit_review = cv.fit(review_texts)
+print('number of unique words in reviews:', len(cv_fit_review.vocabulary_.keys()))
+
+#%% look at common words in summary and review data:
+# get common stop words
+custom_stop_words = set(stopwords.words('english'))
+# add additional stop words:
+ps = PorterStemmer()
+lemmatizer = WordNetLemmatizer() 
+additional_stop_words = {'pop', 'classical', 'jazz', 'dance', 'electronic', 'rock'}
+additional_stop_words_normalized=set()
+for additional_word in additional_stop_words:
+    category_lem = lemmatizer.lemmatize(additional_word)
+    category_stem = ps.stem(additional_word)
+    additional_stop_words_normalized.update({category_lem,category_stem})
+
+custom_stop_words.update(additional_stop_words_normalized)
+summary_texts = list(df['summary'])
+
+
+cv_summary = CountVectorizer(ngram_range=(1,1), lowercase = True, stop_words=list(custom_stop_words), analyzer = 'word', binary = True, max_features = 30)
+cv_summary_fit = cv_summary.fit_transform(summary_texts)
+
+word_count = np.sum(cv_summary_fit.toarray(), axis = 0)
+
+df_temp = pd.DataFrame({'word':cv_summary.get_feature_names(), 'word_count':word_count})
+df_temp.sort_values(by='word_count', ascending=False, inplace = True)
+df_temp.plot(x='word', y='word_count', kind = 'bar')
+plt.xlabel('Top 30 Frequent Unigrams')
+plt.ylabel('Frequency')
+plt.legend('')
+plt.title('Summary Word Distribution')
+plt.ylim(0, 30000)
+plt.show()
+#%%
+review_texts = list(df['reviewText'])
+cv_review = CountVectorizer(ngram_range=(1,1), lowercase = True, stop_words=list(custom_stop_words), analyzer = 'word', binary = True, max_features = 30)
+cv_review_fit = cv_review.fit_transform(review_texts)
+
+word_count = np.sum(cv_review_fit.toarray(), axis = 0)
+
+df_temp = pd.DataFrame({'word':cv_review.get_feature_names(), 'word_count':word_count})
+df_temp.sort_values(by='word_count', ascending=False, inplace = True)
+df_temp.plot(x='word', y='word_count', kind = 'bar')
+plt.xlabel('Top 30 Frequent Unigrams')
+plt.ylabel('Frequency')
+plt.legend('')
+plt.ylim(0, 70000)
+plt.title('Review Word Distribution')
+plt.show()
+
+
+
+
+
+
+
+
+
+
+
+
+
+