#!/usr/bin/env python
# coding: utf-8
#
#
#
1. Business Problem
# 1.1 Problem Description
#
# Netflix is all about connecting people to the movies they love. To help customers find those movies, they developed world-class movie recommendation system: CinematchSM. Its job is to predict whether someone will enjoy a movie based on how much they liked or disliked other movies. Netflix use those predictions to make personal movie recommendations based on each customer’s unique tastes. And while Cinematch is doing pretty well, it can always be made better.
#
# Now there are a lot of interesting alternative approaches to how Cinematch works that netflix haven’t tried. Some are described in the literature, some aren’t. We’re curious whether any of these can beat Cinematch by making better predictions. Because, frankly, if there is a much better approach it could make a big difference to our customers and our business.
# Credits: https://www.netflixprize.com/rules.html
# 1.2 Problem Statement
#
# Netflix provided a lot of anonymous rating data, and a prediction accuracy bar that is 10% better than what Cinematch can do on the same training data set. (Accuracy is a measurement of how closely predicted ratings of movies match subsequent actual ratings.)
#
# 1.3 Sources
#
# - https://www.netflixprize.com/rules.html
# - https://www.kaggle.com/netflix-inc/netflix-prize-data
# - Netflix blog: https://medium.com/netflix-techblog/netflix-recommendations-beyond-the-5-stars-part-1-55838468f429 (very nice blog)
# - surprise library: http://surpriselib.com/ (we use many models from this library)
# - surprise library doc: http://surprise.readthedocs.io/en/stable/getting_started.html (we use many models from this library)
# - installing surprise: https://github.com/NicolasHug/Surprise#installation
# - Research paper: http://courses.ischool.berkeley.edu/i290-dm/s11/SECURE/a1-koren.pdf (most of our work was inspired by this paper)
# - SVD Decomposition : https://www.youtube.com/watch?v=P5mlg91as1c
#
# 1.4 Real world/Business Objectives and constraints
# Objectives:
# 1. Predict the rating that a user would give to a movie that he ahs not yet rated.
# 2. Minimize the difference between predicted and actual rating (RMSE and MAPE)
#
#
# Constraints:
# 1. Some form of interpretability.
# 2. Machine Learning Problem
# 2.1 Data
# 2.1.1 Data Overview
# Get the data from : https://www.kaggle.com/netflix-inc/netflix-prize-data/data
# Data files :
#
# - combined_data_1.txt
# - combined_data_2.txt
# - combined_data_3.txt
# - combined_data_4.txt
# - movie_titles.csv
#
#
# The first line of each file [combined_data_1.txt, combined_data_2.txt, combined_data_3.txt, combined_data_4.txt] contains the movie id followed by a colon. Each subsequent line in the file corresponds to a rating from a customer and its date in the following format:
#
# CustomerID,Rating,Date
#
# MovieIDs range from 1 to 17770 sequentially.
# CustomerIDs range from 1 to 2649429, with gaps. There are 480189 users.
# Ratings are on a five star (integral) scale from 1 to 5.
# Dates have the format YYYY-MM-DD.
#
# 2.1.2 Example Data point
#
# 1:
# 1488844,3,2005-09-06
# 822109,5,2005-05-13
# 885013,4,2005-10-19
# 30878,4,2005-12-26
# 823519,3,2004-05-03
# 893988,3,2005-11-17
# 124105,4,2004-08-05
# 1248029,3,2004-04-22
# 1842128,4,2004-05-09
# 2238063,3,2005-05-11
# 1503895,4,2005-05-19
# 2207774,5,2005-06-06
# 2590061,3,2004-08-12
# 2442,3,2004-04-14
# 543865,4,2004-05-28
# 1209119,4,2004-03-23
# 804919,4,2004-06-10
# 1086807,3,2004-12-28
# 1711859,4,2005-05-08
# 372233,5,2005-11-23
# 1080361,3,2005-03-28
# 1245640,3,2005-12-19
# 558634,4,2004-12-14
# 2165002,4,2004-04-06
# 1181550,3,2004-02-01
# 1227322,4,2004-02-06
# 427928,4,2004-02-26
# 814701,5,2005-09-29
# 808731,4,2005-10-31
# 662870,5,2005-08-24
# 337541,5,2005-03-23
# 786312,3,2004-11-16
# 1133214,4,2004-03-07
# 1537427,4,2004-03-29
# 1209954,5,2005-05-09
# 2381599,3,2005-09-12
# 525356,2,2004-07-11
# 1910569,4,2004-04-12
# 2263586,4,2004-08-20
# 2421815,2,2004-02-26
# 1009622,1,2005-01-19
# 1481961,2,2005-05-24
# 401047,4,2005-06-03
# 2179073,3,2004-08-29
# 1434636,3,2004-05-01
# 93986,5,2005-10-06
# 1308744,5,2005-10-29
# 2647871,4,2005-12-30
# 1905581,5,2005-08-16
# 2508819,3,2004-05-18
# 1578279,1,2005-05-19
# 1159695,4,2005-02-15
# 2588432,3,2005-03-31
# 2423091,3,2005-09-12
# 470232,4,2004-04-08
# 2148699,2,2004-06-05
# 1342007,3,2004-07-16
# 466135,4,2004-07-13
# 2472440,3,2005-08-13
# 1283744,3,2004-04-17
# 1927580,4,2004-11-08
# 716874,5,2005-05-06
# 4326,4,2005-10-29
#
# 2.2 Mapping the real world problem to a Machine Learning Problem
# 2.2.1 Type of Machine Learning Problem
#
# For a given movie and user we need to predict the rating would be given by him/her to the movie.
# The given problem is a Recommendation problem
# It can also seen as a Regression problem
#
# 2.2.2 Performance metric
#
# - Mean Absolute Percentage Error: https://en.wikipedia.org/wiki/Mean_absolute_percentage_error
# - Root Mean Square Error: https://en.wikipedia.org/wiki/Root-mean-square_deviation
#
#
# 2.2.3 Machine Learning Objective and Constraints
# 1. Minimize RMSE.
# 2. Try to provide some interpretability.
# In[ ]:
# this is just to know how much time will it take to run this entire ipython notebook
from datetime import datetime
# globalstart = datetime.now()
import pandas as pd
import numpy as np
import matplotlib
matplotlib.use('nbagg')
# !pip install scikit-surprise
import matplotlib.pyplot as plt
plt.rcParams.update({'figure.max_open_warning': 0})
import seaborn as sns
sns.set_style('whitegrid')
import os
from scipy import sparse
from scipy.sparse import csr_matrix
get_ipython().run_line_magic('matplotlib', 'inline')
from sklearn.decomposition import TruncatedSVD
from sklearn.metrics.pairwise import cosine_similarity
import random
# In[ ]:
# In[ ]:
from google.colab import drive
drive.mount('/content/drive')
#
# 3. Exploratory Data Analysis
# 3.1 Preprocessing
# 3.1.1 Converting / Merging whole data to required format: u_i, m_j, r_ij
# In[ ]:
start = datetime.now()
if not os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/data.csv'):
# Create a file 'data.csv' before reading it
# Read all the files in netflix and store them in one big file('data.csv')
# We re reading from each of the four files and appendig each rating to a global file 'train.csv'
data = open('/content/drive/My Drive/Netflix_recommender/data_folder/data.csv', mode='w')
row = list()
files=['/content/drive/My Drive/Netflix_recommender/data_folder/combined_data_1.txt','/content/drive/My Drive/Netflix_recommender/data_folder/combined_data_2.txt',
'/content/drive/My Drive/Netflix_recommender/data_folder/combined_data_3.txt', '/content/drive/My Drive/Netflix_recommender/data_folder/combined_data_4.txt']
for file in files:
print("Reading ratings from {}...".format(file))
with open(file) as f:
for line in f:
del row[:] # you don't have to do this.
line = line.strip()
if line.endswith(':'):
# All below are ratings for this movie, until another movie appears.
movie_id = line.replace(':', '')
else:
row = [x for x in line.split(',')]
row.insert(0, movie_id)
data.write(','.join(row))
data.write('\n')
print("Done.\n")
data.close()
print('Time taken :', datetime.now() - start)
#
# In[ ]:
print("creating the dataframe from data.csv file..")
df = pd.read_csv('/content/drive/My Drive/Netflix_recommender/data_folder/data.csv', sep=',',
names=['movie', 'user','rating','date'])
df.date = pd.to_datetime(df.date)
print('Done.\n')
# we are arranging the ratings according to time.
print('Sorting the dataframe by date..')
df.sort_values(by='date', inplace=True)
print('Done..')
# In[ ]:
df.head()
# In[ ]:
df.describe()['rating']
# 3.1.2 Checking for NaN values
# In[ ]:
# just to make sure that all Nan containing rows are deleted..
print("No of Nan values in our dataframe : ", sum(df.isnull().any()))
#
# 3.1.3 Removing Duplicates
# In[ ]:
dup_bool = df.duplicated(['movie','user','rating'])
dups = sum(dup_bool) # by considering all columns..( including timestamp)
print("There are {} duplicate rating entries in the data..".format(dups))
#
# 3.1.4 Basic Statistics (#Ratings, #Users, and #Movies)
# In[ ]:
print("Total data ")
print("-"*50)
print("\nTotal no of ratings :",df.shape[0])
print("Total No of Users :", len(np.unique(df.user)))
print("Total No of movies :", len(np.unique(df.movie)))
# 3.2 Spliting data into Train and Test(80:20)
# In[ ]:
if not os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/train.csv'):
# create the dataframe and store it in the disk for offline purposes..
df.iloc[:int(df.shape[0]*0.80)].to_csv("train.csv", index=False)
if not os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/test.csv'):
# create the dataframe and store it in the disk for offline purposes..
df.iloc[int(df.shape[0]*0.80):].to_csv("test.csv", index=False)
train_df = pd.read_csv("/content/drive/My Drive/Netflix_recommender/data_folder/train.csv", parse_dates=['date'])
test_df = pd.read_csv("/content/drive/My Drive/Netflix_recommender/data_folder/test.csv")
# 3.2.1 Basic Statistics in Train data (#Ratings, #Users, and #Movies)
# In[ ]:
# movies = train_df.movie.value_counts()
# users = train_df.user.value_counts()
print("Training data ")
print("-"*50)
print("\nTotal no of ratings :",train_df.shape[0])
print("Total No of Users :", len(np.unique(train_df.user)))
print("Total No of movies :", len(np.unique(train_df.movie)))
# 3.2.2 Basic Statistics in Test data (#Ratings, #Users, and #Movies)
# In[ ]:
print("Test data ")
print("-"*50)
print("\nTotal no of ratings :",test_df.shape[0])
print("Total No of Users :", len(np.unique(test_df.user)))
print("Total No of movies :", len(np.unique(test_df.movie)))
# 3.3 Exploratory Data Analysis on Train data
#
# In[ ]:
# method to make y-axis more readable
def human(num, units = 'M'):
units = units.lower()
num = float(num)
if units == 'k':
return str(num/10**3) + " K"
elif units == 'm':
return str(num/10**6) + " M"
elif units == 'b':
return str(num/10**9) + " B"
# 3.3.1 Distribution of ratings
# In[ ]:
fig, ax = plt.subplots()
plt.title('Distribution of ratings over Training dataset', fontsize=15)
sns.countplot(train_df.rating)
ax.set_yticklabels([human(item, 'M') for item in ax.get_yticks()])
ax.set_ylabel('No. of Ratings(Millions)')
plt.show()
# Add new column (week day) to the data set for analysis.
# In[ ]:
# It is used to skip the warning ''SettingWithCopyWarning''..
pd.options.mode.chained_assignment = None # default='warn'
train_df['day_of_week'] = train_df.date.dt.weekday_name
train_df.tail()
# 3.3.2 Number of Ratings per a month
# In[ ]:
ax = train_df.resample('m', on='date')['rating'].count().plot()
ax.set_title('No of ratings per month (Training data)')
plt.xlabel('Month')
plt.ylabel('No of ratings(per month)')
ax.set_yticklabels([human(item, 'M') for item in ax.get_yticks()])
plt.show()
#
# 3.3.3 Analysis on the Ratings given by user
# In[ ]:
no_of_rated_movies_per_user = train_df.groupby(by='user')['rating'].count().sort_values(ascending=False)
no_of_rated_movies_per_user.head()
# In[ ]:
no_of_rated_movies_per_user.describe()
# In[ ]:
plt.plot(no_of_rated_movies_per_user.values)
# In[ ]:
plt.plot(no_of_rated_movies_per_user.values[:100])
# In[ ]:
tyfig = plt.figure(figsize=plt.figaspect(.5))
ax1 = plt.subplot(121)
sns.kdeplot(no_of_rated_movies_per_user, shade=True, ax=ax1)
plt.xlabel('No of ratings by user')
plt.title("PDF")
ax2 = plt.subplot(122)
sns.kdeplot(no_of_rated_movies_per_user, shade=True, cumulative=True,ax=ax2)
plt.xlabel('No of ratings by user')
plt.title('CDF')
plt.show()
# In[ ]:
no_of_rated_movies_per_user.describe()
# > _There, is something interesting going on with the quantiles.._
# In[ ]:
quantiles = no_of_rated_movies_per_user.quantile(np.arange(0,1.01,0.01), interpolation='higher')
# In[ ]:
# In[ ]:
plt.title("Quantiles and their Values")
quantiles.plot()
# quantiles with 0.05 difference
plt.scatter(x=quantiles.index[::5], y=quantiles.values[::5], c='orange', label="quantiles with 0.05 intervals")
# quantiles with 0.25 difference
plt.scatter(x=quantiles.index[::25], y=quantiles.values[::25], c='m', label = "quantiles with 0.25 intervals")
plt.ylabel('No of ratings by user')
plt.xlabel('Value at the quantile')
plt.legend(loc='best')
# annotate the 25th, 50th, 75th and 100th percentile values....
for x,y in zip(quantiles.index[::25], quantiles[::25]):
plt.annotate(s="({} , {})".format(x,y), xy=(x,y), xytext=(x-0.05, y+500)
,fontweight='bold')
plt.show()
# In[ ]:
quantiles[::5]
# __how many ratings at the last 5% of all ratings__??
# In[ ]:
print('\n No of ratings at last 5 percentile : {}\n'.format(sum(no_of_rated_movies_per_user>= 749)) )
# 3.3.4 Analysis of ratings of a movie given by a user
# In[ ]:
no_of_ratings_per_movie = train_df.groupby(by='movie')['rating'].count().sort_values(ascending=False)
fig = plt.figure(figsize=plt.figaspect(.5))
ax = plt.gca()
plt.plot(no_of_ratings_per_movie.values)
plt.title('# RATINGS per Movie')
plt.xlabel('Movie')
plt.ylabel('No of Users who rated a movie')
ax.set_xticklabels([])
plt.show()
# - __It is very skewed.. just like nunmber of ratings given per user.__
#
#
# - There are some movies (which are very popular) which are rated by huge number of users.
#
# - But most of the movies(like 90%) got some hundereds of ratings.
# 3.3.5 Number of ratings on each day of the week
# In[ ]:
fig, ax = plt.subplots()
sns.countplot(x='day_of_week', data=train_df, ax=ax)
plt.title('No of ratings on each day...')
plt.ylabel('Total no of ratings')
plt.xlabel('')
ax.set_yticklabels([human(item, 'M') for item in ax.get_yticks()])
plt.show()
# In[ ]:
start = datetime.now()
fig = plt.figure(figsize=plt.figaspect(.45))
sns.boxplot(y='rating', x='day_of_week', data=train_df)
plt.show()
print(datetime.now() - start)
# In[ ]:
avg_week_df = train_df.groupby(by=['day_of_week'])['rating'].mean()
print(" AVerage ratings")
print("-"*30)
print(avg_week_df)
print("\n")
#
# 3.3.6 Creating sparse matrix from data frame
#
# 3.3.6.1 Creating sparse matrix from train data frame
# In[ ]:
start = datetime.now()
if os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/train_sparse_matrix.npz'):
print("It is present in your pwd, getting it from disk....")
# just get it from the disk instead of computing it
train_sparse_matrix = sparse.load_npz('/content/drive/My Drive/Netflix_recommender/data_folder/train_sparse_matrix.npz')
print("DONE..")
else:
print("We are creating sparse_matrix from the dataframe..")
# create sparse_matrix and store it for after usage.
# csr_matrix(data_values, (row_index, col_index), shape_of_matrix)
# It should be in such a way that, MATRIX[row, col] = data
train_sparse_matrix = sparse.csr_matrix((train_df.rating.values, (train_df.user.values,
train_df.movie.values)),)
print('Done. It\'s shape is : (user, movie) : ',train_sparse_matrix.shape)
print('Saving it into disk for furthur usage..')
# save it into disk
sparse.save_npz("train_sparse_matrix.npz", train_sparse_matrix)
print('Done..\n')
print(datetime.now() - start)
# The Sparsity of Train Sparse Matrix
# In[ ]:
us,mv = train_sparse_matrix.shape
elem = train_sparse_matrix.count_nonzero()
print("Sparsity Of Train matrix : {} % ".format( (1-(elem/(us*mv))) * 100) )
# 3.3.6.2 Creating sparse matrix from test data frame
# In[ ]:
start = datetime.now()
if os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/test_sparse_matrix.npz'):
print("It is present in your pwd, getting it from disk....")
# just get it from the disk instead of computing it
test_sparse_matrix = sparse.load_npz('/content/drive/My Drive/Netflix_recommender/data_folder/test_sparse_matrix.npz')
print("DONE..")
else:
print("We are creating sparse_matrix from the dataframe..")
# create sparse_matrix and store it for after usage.
# csr_matrix(data_values, (row_index, col_index), shape_of_matrix)
# It should be in such a way that, MATRIX[row, col] = data
test_sparse_matrix = sparse.csr_matrix((test_df.rating.values, (test_df.user.values,
test_df.movie.values)))
print('Done. It\'s shape is : (user, movie) : ',test_sparse_matrix.shape)
print('Saving it into disk for furthur usage..')
# save it into disk
sparse.save_npz("test_sparse_matrix.npz", test_sparse_matrix)
print('Done..\n')
print(datetime.now() - start)
# The Sparsity of Test data Matrix
# In[ ]:
us,mv = test_sparse_matrix.shape
elem = test_sparse_matrix.count_nonzero()
print("Sparsity Of Test matrix : {} % ".format( (1-(elem/(us*mv))) * 100) )
# 3.3.7 Finding Global average of all movie ratings, Average rating per user, and Average rating per movie
# In[ ]:
# get the user averages in dictionary (key: user_id/movie_id, value: avg rating)
def get_average_ratings(sparse_matrix, of_users):
# average ratings of user/axes
ax = 1 if of_users else 0 # 1 - User axes,0 - Movie axes
# ".A1" is for converting Column_Matrix to 1-D numpy array
sum_of_ratings = sparse_matrix.sum(axis=ax).A1
# Boolean matrix of ratings ( whether a user rated that movie or not)
is_rated = sparse_matrix!=0
# no of ratings that each user OR movie..
no_of_ratings = is_rated.sum(axis=ax).A1
# max_user and max_movie ids in sparse matrix
u,m = sparse_matrix.shape
# creae a dictonary of users and their average ratigns..
average_ratings = { i : sum_of_ratings[i]/no_of_ratings[i]
for i in range(u if of_users else m)
if no_of_ratings[i] !=0}
# return that dictionary of average ratings
return average_ratings
# 3.3.7.1 finding global average of all movie ratings
# In[ ]:
train_averages = dict()
# get the global average of ratings in our train set.
train_global_average = train_sparse_matrix.sum()/train_sparse_matrix.count_nonzero()
train_averages['global'] = train_global_average
train_averages
# 3.3.7.2 finding average rating per user
# In[ ]:
train_averages['user'] = get_average_ratings(train_sparse_matrix, of_users=True)
print('\nAverage rating of user 10 :',train_averages['user'][10])
# 3.3.7.3 finding average rating per movie
# In[ ]:
train_averages['movie'] = get_average_ratings(train_sparse_matrix, of_users=False)
print('\n AVerage rating of movie 15 :',train_averages['movie'][15])
#
# 3.3.7.4 PDF's & CDF's of Avg.Ratings of Users & Movies (In Train Data)
# In[ ]:
start = datetime.now()
# draw pdfs for average rating per user and average
fig, (ax1, ax2) = plt.subplots(nrows=1, ncols=2, figsize=plt.figaspect(.5))
fig.suptitle('Avg Ratings per User and per Movie', fontsize=15)
ax1.set_title('Users-Avg-Ratings')
# get the list of average user ratings from the averages dictionary..
user_averages = [rat for rat in train_averages['user'].values()]
sns.distplot(user_averages, ax=ax1, hist=False,
kde_kws=dict(cumulative=True), label='Cdf')
sns.distplot(user_averages, ax=ax1, hist=False,label='Pdf')
ax2.set_title('Movies-Avg-Rating')
# get the list of movie_average_ratings from the dictionary..
movie_averages = [rat for rat in train_averages['movie'].values()]
sns.distplot(movie_averages, ax=ax2, hist=False,
kde_kws=dict(cumulative=True), label='Cdf')
sns.distplot(movie_averages, ax=ax2, hist=False, label='Pdf')
plt.show()
print(datetime.now() - start)
# 3.3.8 Cold Start problem
# 3.3.8.1 Cold Start problem with Users
# In[ ]:
total_users = len(np.unique(df.user))
users_train = len(train_averages['user'])
new_users = total_users - users_train
print('\nTotal number of Users :', total_users)
print('\nNumber of Users in Train data :', users_train)
print("\nNo of Users that didn't appear in train data: {}({} %) \n ".format(new_users,
np.round((new_users/total_users)*100, 2)))
# > We might have to handle __new users__ ( ___75148___ ) who didn't appear in train data.
# 3.3.8.2 Cold Start problem with Movies
# In[ ]:
total_movies = len(np.unique(df.movie))
movies_train = len(train_averages['movie'])
new_movies = total_movies - movies_train
print('\nTotal number of Movies :', total_movies)
print('\nNumber of Users in Train data :', movies_train)
print("\nNo of Movies that didn't appear in train data: {}({} %) \n ".format(new_movies,
np.round((new_movies/total_movies)*100, 2)))
# > We might have to handle __346 movies__ (small comparatively) in test data
#
# 3.4 Computing Similarity matrices
# 3.4.1 Computing User-User Similarity matrix
# 1. Calculating User User Similarity_Matrix is __not very easy__(_unless you have huge Computing Power and lots of time_) because of number of. usersbeing lare.
#
# * You can try if you want to. Your system could crash or the program stops with **Memory Error**
#
# 3.4.1.1 Trying with all dimensions (17k dimensions per user)
# In[ ]:
from sklearn.metrics.pairwise import cosine_similarity
def compute_user_similarity(sparse_matrix, compute_for_few=False, top = 100, verbose=False, verb_for_n_rows = 20,
draw_time_taken=True):
no_of_users, _ = sparse_matrix.shape
# get the indices of non zero rows(users) from our sparse matrix
row_ind, col_ind = sparse_matrix.nonzero()
row_ind = sorted(set(row_ind)) # we don't have to
time_taken = list() # time taken for finding similar users for an user..
# we create rows, cols, and data lists.., which can be used to create sparse matrices
rows, cols, data = list(), list(), list()
if verbose: print("Computing top",top,"similarities for each user..")
start = datetime.now()
temp = 0
for row in row_ind[:top] if compute_for_few else row_ind:
temp = temp+1
prev = datetime.now()
# get the similarity row for this user with all other users
sim = cosine_similarity(sparse_matrix.getrow(row), sparse_matrix).ravel()
# We will get only the top ''top'' most similar users and ignore rest of them..
top_sim_ind = sim.argsort()[-top:]
top_sim_val = sim[top_sim_ind]
# add them to our rows, cols and data
rows.extend([row]*top)
cols.extend(top_sim_ind)
data.extend(top_sim_val)
time_taken.append(datetime.now().timestamp() - prev.timestamp())
if verbose:
if temp%verb_for_n_rows == 0:
print("computing done for {} users [ time elapsed : {} ]"
.format(temp, datetime.now()-start))
# lets create sparse matrix out of these and return it
if verbose: print('Creating Sparse matrix from the computed similarities')
#return rows, cols, data
if draw_time_taken:
plt.plot(time_taken, label = 'time taken for each user')
plt.plot(np.cumsum(time_taken), label='Total time')
plt.legend(loc='best')
plt.xlabel('User')
plt.ylabel('Time (seconds)')
plt.show()
return sparse.csr_matrix((data, (rows, cols)), shape=(no_of_users, no_of_users)), time_taken
# In[ ]:
start = datetime.now()
u_u_sim_sparse, _ = compute_user_similarity(train_sparse_matrix, compute_for_few=True, top = 100,
verbose=True)
print("-"*100)
print("Time taken :",datetime.now()-start)
# 3.4.1.2 Trying with reduced dimensions (Using TruncatedSVD for dimensionality reduction of user vector)
#
# * We have **405,041 users** in out training set and computing similarities between them..( **17K dimensional vector..**) is time consuming..
#
#
# - From above plot, It took roughly __8.88 sec__ for computing simlilar users for __one user__
#
#
# - We have __405,041 users__ with us in training set.
#
#
# - ${ 405041 \times 8.88 = 3596764.08 \sec } = 59946.068 \min = 999.101133333 \text{ hours}
# = 41.629213889 \text{ days}...$
#
# - Even if we run on 4 cores parallelly (a typical system now a days), It will still take almost __10 and 1/2__ days.
#
# IDEA: Instead, we will try to reduce the dimentsions using SVD, so that __it might__ speed up the process...
# In[ ]:
from datetime import datetime
from sklearn.decomposition import TruncatedSVD
start = datetime.now()
# initilaize the algorithm with some parameters..
# All of them are default except n_components. n_itr is for Randomized SVD solver.
netflix_svd = TruncatedSVD(n_components=500, algorithm='randomized', random_state=15)
trunc_svd = netflix_svd.fit_transform(train_sparse_matrix)
print(datetime.now()-start)
# Here,
#
#
# - $\sum \longleftarrow$ (netflix\_svd.**singular\_values\_** )
#
#
# - $\bigvee^T \longleftarrow$ (netflix\_svd.**components_**)
#
#
# - $\bigcup$ is not returned. instead **Projection_of_X** onto the new vectorspace is returned.
#
#
# - It uses **randomized svd** internally, which returns **All 3 of them saperately**. Use that instead..
# In[ ]:
expl_var = np.cumsum(netflix_svd.explained_variance_ratio_)
# In[ ]:
fig, (ax1, ax2) = plt.subplots(nrows=1, ncols=2, figsize=plt.figaspect(.5))
ax1.set_ylabel("Variance Explained", fontsize=15)
ax1.set_xlabel("# Latent Facors", fontsize=15)
ax1.plot(expl_var)
# annote some (latentfactors, expl_var) to make it clear
ind = [1, 2,4,8,20, 60, 100, 200, 300, 400, 500]
ax1.scatter(x = [i-1 for i in ind], y = expl_var[[i-1 for i in ind]], c='#ff3300')
for i in ind:
ax1.annotate(s ="({}, {})".format(i, np.round(expl_var[i-1], 2)), xy=(i-1, expl_var[i-1]),
xytext = ( i+20, expl_var[i-1] - 0.01), fontweight='bold')
change_in_expl_var = [expl_var[i+1] - expl_var[i] for i in range(len(expl_var)-1)]
ax2.plot(change_in_expl_var)
ax2.set_ylabel("Gain in Var_Expl with One Additional LF", fontsize=10)
ax2.yaxis.set_label_position("right")
ax2.set_xlabel("# Latent Facors", fontsize=20)
plt.show()
# In[ ]:
for i in ind:
print("({}, {})".format(i, np.round(expl_var[i-1], 2)))
#
# > I think 500 dimensions is good enough
#
# ---------
#
# - By just taking __(20 to 30)__ latent factors, explained variance that we could get is __20 %__.
#
# - To take it to __60%__, we have to take __almost 400 latent factors__. It is not fare.
#
#
#
# - It basically is the __gain of variance explained__, if we ___add one additional latent factor to it.___
#
#
# - By adding one by one latent factore too it, the ___gain in expained variance__ with that addition is decreasing. (Obviously, because they are sorted that way).
# - ___LHS Graph___:
# - __x__ --- ( No of latent factos ),
# - __y__ --- ( The variance explained by taking x latent factors)
#
#
#
# - __More decrease in the line (RHS graph) __:
# - We are getting more expained variance than before.
# - __Less decrease in that line (RHS graph)__ :
# - We are not getting benifitted from adding latent factor furthur. This is what is shown in the plots.
#
#
# - ___RHS Graph___:
# - __x__ --- ( No of latent factors ),
# - __y__ --- ( Gain n Expl_Var by taking one additional latent factor)
# In[ ]:
# Let's project our Original U_M matrix into into 500 Dimensional space...
start = datetime.now()
trunc_matrix = train_sparse_matrix.dot(netflix_svd.components_.T)
print(datetime.now()- start)
# In[ ]:
type(trunc_matrix), trunc_matrix.shape
# * Let's convert this to actual sparse matrix and store it for future purposes
# In[ ]:
if not os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/trunc_sparse_matrix.npz'):
# create that sparse sparse matrix
trunc_sparse_matrix = sparse.csr_matrix(trunc_matrix)
# Save this truncated sparse matrix for later usage..
sparse.save_npz('trunc_sparse_matrix', trunc_sparse_matrix)
else:
trunc_sparse_matrix = sparse.load_npz('/content/drive/My Drive/Netflix_recommender/data_folder/trunc_sparse_matrix.npz')
# In[ ]:
trunc_sparse_matrix.shape
# In[ ]:
start = datetime.now()
trunc_u_u_sim_matrix, _ = compute_user_similarity(trunc_sparse_matrix, compute_for_few=True, top=50, verbose=True,
verb_for_n_rows=10)
print("-"*50)
print("time:",datetime.now()-start)
# **: This is taking more time for each user than Original one.**
# - from above plot, It took almost __12.18__ for computing simlilar users for __one user__
#
#
# - We have __405041 users__ with us in training set.
#
#
# - ${ 405041 \times 12.18 ==== 4933399.38 \sec } ==== 82223.323 \min ==== 1370.388716667 \text{ hours}
# ==== 57.099529861 \text{ days}...$
#
# - Even we run on 4 cores parallelly (a typical system now a days), It will still take almost __(14 - 15) __ days.
#
# - __Why did this happen...??__
#
#
# - Just think about it. It's not that difficult.
#
# ---------------------------------_( sparse & dense..................get it ?? )_-----------------------------------
# __Is there any other way to compute user user similarity..??__
# -An alternative is to compute similar users for a particular user, whenenver required (**ie., Run time**)
# - We maintain a binary Vector for users, which tells us whether we already computed or not..
# - ***If not*** :
# - Compute top (let's just say, 1000) most similar users for this given user, and add this to our datastructure, so that we can just access it(similar users) without recomputing it again.
# -
# - ***If It is already Computed***:
# - Just get it directly from our datastructure, which has that information.
# - In production time, We might have to recompute similarities, if it is computed a long time ago. Because user preferences changes over time. If we could maintain some kind of Timer, which when expires, we have to update it ( recompute it ).
# -
# - ***Which datastructure to use:***
# - It is purely implementation dependant.
# - One simple method is to maintain a **Dictionary Of Dictionaries**.
# -
# - **key :** _userid_
# - __value__: _Again a dictionary_
# - __key__ : _Similar User_
# - __value__: _Similarity Value_
# 3.4.2 Computing Movie-Movie Similarity matrix
# In[ ]:
start = datetime.now()
if not os.path.isfile('/content/drive/My Drive/Netflix_recommender/data_folder/m_m_sim_sparse.npz'):
print("It seems you don't have that file. Computing movie_movie similarity...")
start = datetime.now()
m_m_sim_sparse = cosine_similarity(X=train_sparse_matrix.T, dense_output=False)
print("Done..")
# store this sparse matrix in disk before using it. For future purposes.
print("Saving it to disk without the need of re-computing it again.. ")
sparse.save_npz("m_m_sim_sparse.npz", m_m_sim_sparse)
print("Done..")
else:
print("It is there, We will get it.")
m_m_sim_sparse = sparse.load_npz("/content/drive/My Drive/Netflix_recommender/data_folder/m_m_sim_sparse.npz")
print("Done ...")
print("It's a ",m_m_sim_sparse.shape," dimensional matrix")
print(datetime.now() - start)
# In[ ]:
m_m_sim_sparse.shape
# - Even though we have similarity measure of each movie, with all other movies, We generally don't care much about least similar movies.
#
#
# - Most of the times, only top_xxx similar items matters. It may be 10 or 100.
#
#
# - We take only those top similar movie ratings and store them in a saperate dictionary.
# In[ ]:
movie_ids = np.unique(m_m_sim_sparse.nonzero()[1])
# In[ ]:
start = datetime.now()
similar_movies = dict()
for movie in movie_ids:
# get the top similar movies and store them in the dictionary
sim_movies = m_m_sim_sparse[movie].toarray().ravel().argsort()[::-1][1:]
similar_movies[movie] = sim_movies[:100]
print(datetime.now() - start)
# just testing similar movies for movie_15
similar_movies[15]
#
# 3.4.3 Finding most similar movies using similarity matrix
# __ Does Similarity really works as the way we expected...? __
# _Let's pick some random movie and check for its similar movies...._
# In[ ]:
# First Let's load the movie details into soe dataframe..
# movie details are in 'netflix/movie_titles.csv'
movie_titles = pd.read_csv("data_folder/movie_titles.csv", sep=',', header = None,
names=['movie_id', 'year_of_release', 'title'], verbose=True,
index_col = 'movie_id', encoding = "ISO-8859-1")
movie_titles.head()
# Similar Movies for 'Vampire Journals'
# In[ ]:
mv_id = 67
print("\nMovie ----->",movie_titles.loc[mv_id].values[1])
print("\nIt has {} Ratings from users.".format(train_sparse_matrix[:,mv_id].getnnz()))
print("\nWe have {} movies which are similarto this and we will get only top most..".format(m_m_sim_sparse[:,mv_id].getnnz()))
# In[ ]:
similarities = m_m_sim_sparse[mv_id].toarray().ravel()
similar_indices = similarities.argsort()[::-1][1:]
similarities[similar_indices]
sim_indices = similarities.argsort()[::-1][1:] # It will sort and reverse the array and ignore its similarity (ie.,1)
# and return its indices(movie_ids)
# In[ ]:
plt.plot(similarities[sim_indices], label='All the ratings')
plt.plot(similarities[sim_indices[:100]], label='top 100 similar movies')
plt.title("Similar Movies of {}(movie_id)".format(mv_id), fontsize=20)
plt.xlabel("Movies (Not Movie_Ids)", fontsize=15)
plt.ylabel("Cosine Similarity",fontsize=15)
plt.legend()
plt.show()
#
#
# __Top 10 similar movies__
# In[ ]:
movie_titles.loc[sim_indices[:10]]
#
# > Similarly, we can ___find similar users___ and compare how similar they are.
#
#
# 4. Machine Learning Models
#
# In[ ]:
def get_sample_sparse_matrix(sparse_matrix, no_users, no_movies, path, verbose = True):
"""
It will get it from the ''path'' if it is present or It will create
and store the sampled sparse matrix in the path specified.
"""
# get (row, col) and (rating) tuple from sparse_matrix...
row_ind, col_ind, ratings = sparse.find(sparse_matrix)
users = np.unique(row_ind)
movies = np.unique(col_ind)
print("Original Matrix : (users, movies) -- ({} {})".format(len(users), len(movies)))
print("Original Matrix : Ratings -- {}\n".format(len(ratings)))
# It just to make sure to get same sample everytime we run this program..
# and pick without replacement....
np.random.seed(15)
sample_users = np.random.choice(users, no_users, replace=False)
sample_movies = np.random.choice(movies, no_movies, replace=False)
# get the boolean mask or these sampled_items in originl row/col_inds..
mask = np.logical_and( np.isin(row_ind, sample_users),
np.isin(col_ind, sample_movies) )
sample_sparse_matrix = sparse.csr_matrix((ratings[mask], (row_ind[mask], col_ind[mask])),
shape=(max(sample_users)+1, max(sample_movies)+1))
if verbose:
print("Sampled Matrix : (users, movies) -- ({} {})".format(len(sample_users), len(sample_movies)))
print("Sampled Matrix : Ratings --", format(ratings[mask].shape[0]))
print('Saving it into disk for furthur usage..')
# save it into disk
sparse.save_npz(path, sample_sparse_matrix)
if verbose:
print('Done..\n')
return sample_sparse_matrix
# 4.1 Sampling Data
# 4.1.1 Build sample train data from the train data
# In[ ]:
start = datetime.now()
path = "/content/drive/My Drive/Netflix_recommender/sample_train_sparse_matrix.npz"
if os.path.isfile(path):
print("It is present in your pwd, getting it from disk....")
# just get it from the disk instead of computing it
sample_train_sparse_matrix = sparse.load_npz(path)
print("DONE..")
else:
# get 10k users and 1k movies from available data
sample_train_sparse_matrix = get_sample_sparse_matrix(train_sparse_matrix, no_users=13000, no_movies=1500,
path = path)
print(datetime.now() - start)
# 4.1.2 Build sample test data from the test data
# In[ ]:
start = datetime.now()
path = "/content/drive/My Drive/Netflix_recommender/sample_test_sparse_matrix.npz"
if os.path.isfile(path):
print("It is present in your pwd, getting it from disk....")
# just get it from the disk instead of computing it
sample_test_sparse_matrix = sparse.load_npz(path)
print("DONE..")
else:
# get 5k users and 500 movies from available data
sample_test_sparse_matrix = get_sample_sparse_matrix(test_sparse_matrix, no_users=5000, no_movies=500,
path = "/content/drive/My Drive/Netflix_recommender/sample_test_sparse_matrix.npz")
print(datetime.now() - start)
#
# 4.2 Finding Global Average of all movie ratings, Average rating per User, and Average rating per Movie (from sampled train)
# In[ ]:
sample_train_averages = dict()
# 4.2.1 Finding Global Average of all movie ratings
# In[ ]:
# get the global average of ratings in our train set.
global_average = sample_train_sparse_matrix.sum()/sample_train_sparse_matrix.count_nonzero()
sample_train_averages['global'] = global_average
sample_train_averages
# 4.2.2 Finding Average rating per User
# In[ ]:
sample_train_averages['user'] = get_average_ratings(sample_train_sparse_matrix, of_users=True)
print('\nAverage rating of user 1515220 :',sample_train_averages['user'][1515220])
# 4.2.3 Finding Average rating per Movie
# In[ ]:
sample_train_averages['movie'] = get_average_ratings(sample_train_sparse_matrix, of_users=False)
print('\n AVerage rating of movie 15153 :',sample_train_averages['movie'][15153])
#
# 4.3 Featurizing data
# In[ ]:
print('\n No of ratings in Our Sampled train matrix is : {}\n'.format(sample_train_sparse_matrix.count_nonzero()))
print('\n No of ratings in Our Sampled test matrix is : {}\n'.format(sample_test_sparse_matrix.count_nonzero()))
# 4.3.1 Featurizing data for regression problem
# 4.3.1.1 Featurizing train data
# In[ ]:
# get users, movies and ratings from our samples train sparse matrix
sample_train_users, sample_train_movies, sample_train_ratings = sparse.find(sample_train_sparse_matrix)
# In[ ]:
############################################################
# It took me almost 10 hours to prepare this train dataset.#
############################################################
start = datetime.now()
if os.path.isfile('sample/small/reg_train.csv'):
print("File already exists you don't have to prepare again..." )
else:
print('preparing {} tuples for the dataset..\n'.format(len(sample_train_ratings)))
with open('sample/small/reg_train.csv', mode='w') as reg_data_file:
count = 0
for (user, movie, rating) in zip(sample_train_users, sample_train_movies, sample_train_ratings):
st = datetime.now()
# print(user, movie)
#--------------------- Ratings of "movie" by similar users of "user" ---------------------
# compute the similar Users of the "user"
user_sim = cosine_similarity(sample_train_sparse_matrix[user], sample_train_sparse_matrix).ravel()
top_sim_users = user_sim.argsort()[::-1][1:] # we are ignoring 'The User' from its similar users.
# get the ratings of most similar users for this movie
top_ratings = sample_train_sparse_matrix[top_sim_users, movie].toarray().ravel()
# we will make it's length "5" by adding movie averages to .
top_sim_users_ratings = list(top_ratings[top_ratings != 0][:5])
top_sim_users_ratings.extend([sample_train_averages['movie'][movie]]*(5 - len(top_sim_users_ratings)))
# print(top_sim_users_ratings, end=" ")
#--------------------- Ratings by "user" to similar movies of "movie" ---------------------
# compute the similar movies of the "movie"
movie_sim = cosine_similarity(sample_train_sparse_matrix[:,movie].T, sample_train_sparse_matrix.T).ravel()
top_sim_movies = movie_sim.argsort()[::-1][1:] # we are ignoring 'The User' from its similar users.
# get the ratings of most similar movie rated by this user..
top_ratings = sample_train_sparse_matrix[user, top_sim_movies].toarray().ravel()
# we will make it's length "5" by adding user averages to.
top_sim_movies_ratings = list(top_ratings[top_ratings != 0][:5])
top_sim_movies_ratings.extend([sample_train_averages['user'][user]]*(5-len(top_sim_movies_ratings)))
# print(top_sim_movies_ratings, end=" : -- ")
#-----------------prepare the row to be stores in a file-----------------#
row = list()
row.append(user)
row.append(movie)
# Now add the other features to this data...
row.append(sample_train_averages['global']) # first feature
# next 5 features are similar_users "movie" ratings
row.extend(top_sim_users_ratings)
# next 5 features are "user" ratings for similar_movies
row.extend(top_sim_movies_ratings)
# Avg_user rating
row.append(sample_train_averages['user'][user])
# Avg_movie rating
row.append(sample_train_averages['movie'][movie])
# finalley, The actual Rating of this user-movie pair...
row.append(rating)
count = count + 1
# add rows to the file opened..
reg_data_file.write(','.join(map(str, row)))
reg_data_file.write('\n')
if (count)%10000 == 0:
# print(','.join(map(str, row)))
print("Done for {} rows----- {}".format(count, datetime.now() - start))
print(datetime.now() - start)
# __Reading from the file to make a Train_dataframe__
# In[ ]:
reg_train = pd.read_csv('/content/drive/My Drive/Netflix_recommender/reg_train_25.csv', names = ['user', 'movie', 'GAvg', 'sur1', 'sur2', 'sur3', 'sur4', 'sur5','smr1', 'smr2', 'smr3', 'smr4', 'smr5', 'UAvg', 'MAvg', 'rating'], header=None)
reg_train.head()
# -----------------------
#
# - __GAvg__ : Average rating of all the ratings
#
#
# - __Similar users rating of this movie__:
# - sur1, sur2, sur3, sur4, sur5 ( top 5 similar users who rated that movie.. )
#
#
#
# - __Similar movies rated by this user__:
# - smr1, smr2, smr3, smr4, smr5 ( top 5 similar movies rated by this movie.. )
#
#
# - __UAvg__ : User's Average rating
#
#
# - __MAvg__ : Average rating of this movie
#
#
# - __rating__ : Rating of this movie by this user.
#
# -----------------------
#
# 4.3.1.2 Featurizing test data
# In[ ]:
# get users, movies and ratings from the Sampled Test
sample_test_users, sample_test_movies, sample_test_ratings = sparse.find(sample_test_sparse_matrix)
# In[ ]:
sample_train_averages['global']
# In[ ]:
start = datetime.now()
if os.path.isfile('sample/small/reg_test.csv'):
print("It is already created...")
else:
print('preparing {} tuples for the dataset..\n'.format(len(sample_test_ratings)))
with open('sample/small/reg_test.csv', mode='w') as reg_data_file:
count = 0
for (user, movie, rating) in zip(sample_test_users, sample_test_movies, sample_test_ratings):
st = datetime.now()
#--------------------- Ratings of "movie" by similar users of "user" ---------------------
#print(user, movie)
try:
# compute the similar Users of the "user"
user_sim = cosine_similarity(sample_train_sparse_matrix[user], sample_train_sparse_matrix).ravel()
top_sim_users = user_sim.argsort()[::-1][1:] # we are ignoring 'The User' from its similar users.
# get the ratings of most similar users for this movie
top_ratings = sample_train_sparse_matrix[top_sim_users, movie].toarray().ravel()
# we will make it's length "5" by adding movie averages to .
top_sim_users_ratings = list(top_ratings[top_ratings != 0][:5])
top_sim_users_ratings.extend([sample_train_averages['movie'][movie]]*(5 - len(top_sim_users_ratings)))
# print(top_sim_users_ratings, end="--")
except (IndexError, KeyError):
# It is a new User or new Movie or there are no ratings for given user for top similar movies...
########## Cold STart Problem ##########
top_sim_users_ratings.extend([sample_train_averages['global']]*(5 - len(top_sim_users_ratings)))
#print(top_sim_users_ratings)
except:
print(user, movie)
# we just want KeyErrors to be resolved. Not every Exception...
raise
#--------------------- Ratings by "user" to similar movies of "movie" ---------------------
try:
# compute the similar movies of the "movie"
movie_sim = cosine_similarity(sample_train_sparse_matrix[:,movie].T, sample_train_sparse_matrix.T).ravel()
top_sim_movies = movie_sim.argsort()[::-1][1:] # we are ignoring 'The User' from its similar users.
# get the ratings of most similar movie rated by this user..
top_ratings = sample_train_sparse_matrix[user, top_sim_movies].toarray().ravel()
# we will make it's length "5" by adding user averages to.
top_sim_movies_ratings = list(top_ratings[top_ratings != 0][:5])
top_sim_movies_ratings.extend([sample_train_averages['user'][user]]*(5-len(top_sim_movies_ratings)))
#print(top_sim_movies_ratings)
except (IndexError, KeyError):
#print(top_sim_movies_ratings, end=" : -- ")
top_sim_movies_ratings.extend([sample_train_averages['global']]*(5-len(top_sim_movies_ratings)))
#print(top_sim_movies_ratings)
except :
raise
#-----------------prepare the row to be stores in a file-----------------#
row = list()
# add usser and movie name first
row.append(user)
row.append(movie)
row.append(sample_train_averages['global']) # first feature
#print(row)
# next 5 features are similar_users "movie" ratings
row.extend(top_sim_users_ratings)
#print(row)
# next 5 features are "user" ratings for similar_movies
row.extend(top_sim_movies_ratings)
#print(row)
# Avg_user rating
try:
row.append(sample_train_averages['user'][user])
except KeyError:
row.append(sample_train_averages['global'])
except:
raise
#print(row)
# Avg_movie rating
try:
row.append(sample_train_averages['movie'][movie])
except KeyError:
row.append(sample_train_averages['global'])
except:
raise
#print(row)
# finalley, The actual Rating of this user-movie pair...
row.append(rating)
#print(row)
count = count + 1
# add rows to the file opened..
reg_data_file.write(','.join(map(str, row)))
#print(','.join(map(str, row)))
reg_data_file.write('\n')
if (count)%1000 == 0:
#print(','.join(map(str, row)))
print("Done for {} rows----- {}".format(count, datetime.now() - start))
print("",datetime.now() - start)
# __Reading from the file to make a test dataframe __
# In[ ]:
reg_test_df = pd.read_csv('/content/drive/My Drive/Netflix_recommender/reg_test_25.csv', names = ['user', 'movie', 'GAvg', 'sur1', 'sur2', 'sur3', 'sur4', 'sur5',
'smr1', 'smr2', 'smr3', 'smr4', 'smr5',
'UAvg', 'MAvg', 'rating'], header=None)
reg_test_df.head(4)
# -----------------------
#
# - __GAvg__ : Average rating of all the ratings
#
#
# - __Similar users rating of this movie__:
# - sur1, sur2, sur3, sur4, sur5 ( top 5 simiular users who rated that movie.. )
#
#
#
# - __Similar movies rated by this user__:
# - smr1, smr2, smr3, smr4, smr5 ( top 5 simiular movies rated by this movie.. )
#
#
# - __UAvg__ : User AVerage rating
#
#
# - __MAvg__ : Average rating of this movie
#
#
# - __rating__ : Rating of this movie by this user.
#
# -----------------------
#
# 4.3.2 Transforming data for Surprise models
# In[ ]:
from surprise import Reader, Dataset
# 4.3.2.1 Transforming train data
# - We can't give raw data (movie, user, rating) to train the model in Surprise library.
#
#
# - They have a saperate format for TRAIN and TEST data, which will be useful for training the models like SVD, KNNBaseLineOnly....etc..,in Surprise.
#
#
# - We can form the trainset from a file, or from a Pandas DataFrame.
# http://surprise.readthedocs.io/en/stable/getting_started.html#load-dom-dataframe-py
# In[ ]:
# It is to specify how to read the dataframe.
# for our dataframe, we don't have to specify anything extra..
reader = Reader(rating_scale=(1,5))
# create the traindata from the dataframe...
train_data = Dataset.load_from_df(reg_train[['user', 'movie', 'rating']], reader)
# build the trainset from traindata.., It is of dataset format from surprise library..
trainset = train_data.build_full_trainset()
# 4.3.2.2 Transforming test data
# - Testset is just a list of (user, movie, rating) tuples. (Order in the tuple is impotant)
# In[ ]:
testset = list(zip(reg_test_df.user.values, reg_test_df.movie.values, reg_test_df.rating.values))
testset[:3]
# 4.4 Applying Machine Learning models
#
# - Global dictionary that stores rmse and mape for all the models....
#
# - It stores the metrics in a dictionary of dictionaries
#
# > __keys__ : model names(string)
#
# > __value__: dict(__key__ : metric, __value__ : value )
# In[ ]:
models_evaluation_train = dict()
models_evaluation_test = dict()
models_evaluation_train, models_evaluation_test
#
# > __Utility functions for running regression models__
# In[ ]:
# to get rmse and mape given actual and predicted ratings..
def get_error_metrics(y_true, y_pred):
rmse = np.sqrt(np.mean([ (y_true[i] - y_pred[i])**2 for i in range(len(y_pred)) ]))
mape = np.mean(np.abs( (y_true - y_pred)/y_true )) * 100
return rmse, mape
###################################################################
###################################################################
def run_xgboost(algo, x_train, y_train, x_test, y_test, verbose=True):
"""
It will return train_results and test_results
"""
# dictionaries for storing train and test results
train_results = dict()
test_results = dict()
# fit the model
print('Training the model..')
start =datetime.now()
algo.fit(x_train, y_train, eval_metric = 'rmse')
print('Done. Time taken : {}\n'.format(datetime.now()-start))
print('Done \n')
# from the trained model, get the predictions....
print('Evaluating the model with TRAIN data...')
start =datetime.now()
y_train_pred = algo.predict(x_train)
# get the rmse and mape of train data...
rmse_train, mape_train = get_error_metrics(y_train.values, y_train_pred)
# store the results in train_results dictionary..
train_results = {'rmse': rmse_train,
'mape' : mape_train,
'predictions' : y_train_pred}
#######################################
# get the test data predictions and compute rmse and mape
print('Evaluating Test data')
y_test_pred = algo.predict(x_test)
rmse_test, mape_test = get_error_metrics(y_true=y_test.values, y_pred=y_test_pred)
# store them in our test results dictionary.
test_results = {'rmse': rmse_test,
'mape' : mape_test,
'predictions':y_test_pred}
if verbose:
print('\nTEST DATA')
print('-'*30)
print('RMSE : ', rmse_test)
print('MAPE : ', mape_test)
# return these train and test results...
return train_results, test_results
# > __Utility functions for Surprise modes__
# In[ ]:
# it is just to makesure that all of our algorithms should produce same results
# everytime they run...
my_seed = 15
random.seed(my_seed)
np.random.seed(my_seed)
##########################################################
# get (actual_list , predicted_list) ratings given list
# of predictions (prediction is a class in Surprise).
##########################################################
def get_ratings(predictions):
actual = np.array([pred.r_ui for pred in predictions])
pred = np.array([pred.est for pred in predictions])
return actual, pred
################################################################
# get ''rmse'' and ''mape'' , given list of prediction objecs
################################################################
def get_errors(predictions, print_them=False):
actual, pred = get_ratings(predictions)
rmse = np.sqrt(np.mean((pred - actual)**2))
mape = np.mean(np.abs(pred - actual)/actual)
return rmse, mape*100
##################################################################################
# It will return predicted ratings, rmse and mape of both train and test data #
##################################################################################
def run_surprise(algo, trainset, testset, verbose=True):
'''
return train_dict, test_dict
It returns two dictionaries, one for train and the other is for test
Each of them have 3 key-value pairs, which specify ''rmse'', ''mape'', and ''predicted ratings''.
'''
start = datetime.now()
# dictionaries that stores metrics for train and test..
train = dict()
test = dict()
# train the algorithm with the trainset
st = datetime.now()
print('Training the model...')
algo.fit(trainset)
print('Done. time taken : {} \n'.format(datetime.now()-st))
# ---------------- Evaluating train data--------------------#
st = datetime.now()
print('Evaluating the model with train data..')
# get the train predictions (list of prediction class inside Surprise)
train_preds = algo.test(trainset.build_testset())
# get predicted ratings from the train predictions..
train_actual_ratings, train_pred_ratings = get_ratings(train_preds)
# get ''rmse'' and ''mape'' from the train predictions.
train_rmse, train_mape = get_errors(train_preds)
print('time taken : {}'.format(datetime.now()-st))
if verbose:
print('-'*15)
print('Train Data')
print('-'*15)
print("RMSE : {}\n\nMAPE : {}\n".format(train_rmse, train_mape))
#store them in the train dictionary
if verbose:
print('adding train results in the dictionary..')
train['rmse'] = train_rmse
train['mape'] = train_mape
train['predictions'] = train_pred_ratings
#------------ Evaluating Test data---------------#
st = datetime.now()
print('\nEvaluating for test data...')
# get the predictions( list of prediction classes) of test data
test_preds = algo.test(testset)
# get the predicted ratings from the list of predictions
test_actual_ratings, test_pred_ratings = get_ratings(test_preds)
# get error metrics from the predicted and actual ratings
test_rmse, test_mape = get_errors(test_preds)
print('time taken : {}'.format(datetime.now()-st))
if verbose:
print('-'*15)
print('Test Data')
print('-'*15)
print("RMSE : {}\n\nMAPE : {}\n".format(test_rmse, test_mape))
# store them in test dictionary
if verbose:
print('storing the test results in test dictionary...')
test['rmse'] = test_rmse
test['mape'] = test_mape
test['predictions'] = test_pred_ratings
print('\n'+'-'*45)
print('Total time taken to run this algorithm :', datetime.now() - start)
# return two dictionaries train and test
return train, test
#
# 4.4.1 XGBoost with initial 13 features
# In[ ]:
import xgboost as xgb
# In[ ]:
# prepare Train data
x_train = reg_train.drop(['user','movie','rating'], axis=1)
y_train = reg_train['rating']
# Prepare Test data
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# initialize Our first XGBoost model...
first_xgb = xgb.XGBRegressor(silent=False, n_jobs=13, random_state=15, n_estimators=100)
train_results, test_results = run_xgboost(first_xgb, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
models_evaluation_train['first_algo'] = train_results
models_evaluation_test['first_algo'] = test_results
xgb.plot_importance(first_xgb)
plt.show()
#
# 4.4.2 Suprise BaselineModel
#
#
# In[ ]:
from surprise import BaselineOnly
# __Predicted_rating : ( baseline prediction ) __
#
# - http://surprise.readthedocs.io/en/stable/basic_algorithms.html#surprise.prediction_algorithms.baseline_only.BaselineOnly
# >$ \large {\hat{r}_{ui} = b_{ui} =\mu + b_u + b_i} $
#
#
# - $\pmb \mu $ : Average of all trainings in training data.
# - $\pmb b_u$ : User bias
# - $\pmb b_i$ : Item bias (movie biases)
# __Optimization function ( Least Squares Problem ) __
#
# - http://surprise.readthedocs.io/en/stable/prediction_algorithms.html#baselines-estimates-configuration
#
# > $ \large \sum_{r_{ui} \in R_{train}} \left(r_{ui} - (\mu + b_u + b_i)\right)^2 +
# \lambda \left(b_u^2 + b_i^2 \right).\text { [mimimize } {b_u, b_i]}$
# In[ ]:
# options are to specify.., how to compute those user and item biases
bsl_options = {'method': 'sgd',
'learning_rate': .001
}
bsl_algo = BaselineOnly(bsl_options=bsl_options)
# run this algorithm.., It will return the train and test results..
bsl_train_results, bsl_test_results = run_surprise(my_bsl_algo, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
models_evaluation_train['bsl_algo'] = bsl_train_results
models_evaluation_test['bsl_algo'] = bsl_test_results
#
# 4.4.3 XGBoost with initial 13 features + Surprise Baseline predictor
# __Updating Train Data__
# In[ ]:
# add our baseline_predicted value as our feature..
reg_train['bslpr'] = models_evaluation_train['bsl_algo']['predictions']
reg_train.head(2)
# __Updating Test Data__
# In[ ]:
# add that baseline predicted ratings with Surprise to the test data as well
reg_test_df['bslpr'] = models_evaluation_test['bsl_algo']['predictions']
reg_test_df.head(2)
# In[ ]:
# prepare train data
x_train = reg_train.drop(['user', 'movie','rating'], axis=1)
y_train = reg_train['rating']
# Prepare Test data
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# initialize Our first XGBoost model...
xgb_bsl = xgb.XGBRegressor(silent=False, n_jobs=13, random_state=15, n_estimators=100)
train_results, test_results = run_xgboost(xgb_bsl, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
models_evaluation_train['xgb_bsl'] = train_results
models_evaluation_test['xgb_bsl'] = test_results
xgb.plot_importance(xgb_bsl)
plt.show()
#
#
# 4.4.4 Surprise KNNBaseline predictor
# In[ ]:
from surprise import KNNBaseline
# - KNN BASELINE
# - http://surprise.readthedocs.io/en/stable/knn_inspired.html#surprise.prediction_algorithms.knns.KNNBaseline
# - PEARSON_BASELINE SIMILARITY
# - http://surprise.readthedocs.io/en/stable/similarities.html#surprise.similarities.pearson_baseline
# - SHRINKAGE
# - _2.2 Neighborhood Models_ in http://courses.ischool.berkeley.edu/i290-dm/s11/SECURE/a1-koren.pdf
# - __predicted Rating__ : ( ___ based on User-User similarity ___ )
#
# \begin{align} \hat{r}_{ui} = b_{ui} + \frac{ \sum\limits_{v \in N^k_i(u)}
# \text{sim}(u, v) \cdot (r_{vi} - b_{vi})} {\sum\limits_{v \in
# N^k_i(u)} \text{sim}(u, v)} \end{align}
#
# - $\pmb{b_{ui}}$ - _Baseline prediction_ of (user,movie) rating
#
# - $ \pmb {N_i^k (u)}$ - Set of __K similar__ users (neighbours) of __user (u)__ who rated __movie(i)__
#
# - _sim (u, v)_ - __Similarity__ between users __u and v__
# - Generally, it will be cosine similarity or Pearson correlation coefficient.
# - But we use __shrunk Pearson-baseline correlation coefficient__, which is based on the pearsonBaseline similarity ( we take base line predictions instead of mean rating of user/item)
#
#
# - __ Predicted rating __ ( based on Item Item similarity ):
# \begin{align} \hat{r}_{ui} = b_{ui} + \frac{ \sum\limits_{j \in N^k_u(i)}\text{sim}(i, j) \cdot (r_{uj} - b_{uj})} {\sum\limits_{j \in N^k_u(j)} \text{sim}(i, j)} \end{align}
#
# - ___Notations follows same as above (user user based predicted rating ) ___
# 4.4.4.1 Surprise KNNBaseline with user user similarities
# In[ ]:
# we specify , how to compute similarities and what to consider with sim_options to our algorithm
sim_options = {'user_based' : True,
'name': 'pearson_baseline',
'shrinkage': 100,
'min_support': 2
}
# we keep other parameters like regularization parameter and learning_rate as default values.
bsl_options = {'method': 'sgd'}
knn_bsl_u = KNNBaseline(k=40, sim_options = sim_options, bsl_options = bsl_options)
knn_bsl_u_train_results, knn_bsl_u_test_results = run_surprise(knn_bsl_u, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
models_evaluation_train['knn_bsl_u'] = knn_bsl_u_train_results
models_evaluation_test['knn_bsl_u'] = knn_bsl_u_test_results
# 4.4.4.2 Surprise KNNBaseline with movie movie similarities
# In[ ]:
# we specify , how to compute similarities and what to consider with sim_options to our algorithm
# 'user_based' : Fals => this considers the similarities of movies instead of users
sim_options = {'user_based' : False,
'name': 'pearson_baseline',
'shrinkage': 100,
'min_support': 2
}
# we keep other parameters like regularization parameter and learning_rate as default values.
bsl_options = {'method': 'sgd'}
knn_bsl_m = KNNBaseline(k=40, sim_options = sim_options, bsl_options = bsl_options)
knn_bsl_m_train_results, knn_bsl_m_test_results = run_surprise(knn_bsl_m, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
models_evaluation_train['knn_bsl_m'] = knn_bsl_m_train_results
models_evaluation_test['knn_bsl_m'] = knn_bsl_m_test_results
#
# 4.4.5 XGBoost with initial 13 features + Surprise Baseline predictor + KNNBaseline predictor
# - - - First we will run XGBoost with predictions from both KNN's ( that uses User\_User and Item\_Item similarities along with our previous features.
#
#
# - - - Then we will run XGBoost with just predictions form both knn models and preditions from our baseline model.
# __Preparing Train data __
# In[ ]:
# add the predicted values from both knns to this dataframe
reg_train['knn_bsl_u'] = models_evaluation_train['knn_bsl_u']['predictions']
reg_train['knn_bsl_m'] = models_evaluation_train['knn_bsl_m']['predictions']
reg_train.head(2)
# __Preparing Test data __
# In[ ]:
reg_test_df['knn_bsl_u'] = models_evaluation_test['knn_bsl_u']['predictions']
reg_test_df['knn_bsl_m'] = models_evaluation_test['knn_bsl_m']['predictions']
reg_test_df.head(2)
# In[ ]:
# prepare the train data....
x_train = reg_train.drop(['user', 'movie', 'rating'], axis=1)
y_train = reg_train['rating']
# prepare the train data....
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# declare the model
xgb_knn_bsl = xgb.XGBRegressor(n_jobs=10, random_state=15)
train_results, test_results = run_xgboost(xgb_knn_bsl, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
models_evaluation_train['xgb_knn_bsl'] = train_results
models_evaluation_test['xgb_knn_bsl'] = test_results
xgb.plot_importance(xgb_knn_bsl)
plt.show()
# 4.4.6 Matrix Factorization Techniques
# 4.4.6.1 SVD Matrix Factorization User Movie intractions
# In[ ]:
from surprise import SVD
# http://surprise.readthedocs.io/en/stable/matrix_factorization.html#surprise.prediction_algorithms.matrix_factorization.SVD
# - __ Predicted Rating : __
# -
# - $ \large \hat r_{ui} = \mu + b_u + b_i + q_i^Tp_u $
#
# - $\pmb q_i$ - Representation of item(movie) in latent factor space
#
# - $\pmb p_u$ - Representation of user in new latent factor space
#
#
# - A BASIC MATRIX FACTORIZATION MODEL in https://datajobs.com/data-science-repo/Recommender-Systems-[Netflix].pdf
# - __Optimization problem with user item interactions and regularization (to avoid overfitting)__
# -
# - $\large \sum_{r_{ui} \in R_{train}} \left(r_{ui} - \hat{r}_{ui} \right)^2 +
# \lambda\left(b_i^2 + b_u^2 + ||q_i||^2 + ||p_u||^2\right) $
# In[ ]:
# initiallize the model
svd = SVD(n_factors=100, biased=True, random_state=15, verbose=True)
svd_train_results, svd_test_results = run_surprise(svd, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
models_evaluation_train['svd'] = svd_train_results
models_evaluation_test['svd'] = svd_test_results
#
# 4.4.6.2 SVD Matrix Factorization with implicit feedback from user ( user rated movies )
# In[ ]:
from surprise import SVDpp
# - -----> 2.5 Implicit Feedback in http://courses.ischool.berkeley.edu/i290-dm/s11/SECURE/a1-koren.pdf
# - __ Predicted Rating : __
# -
# - $ \large \hat{r}_{ui} = \mu + b_u + b_i + q_i^T\left(p_u +
# |I_u|^{-\frac{1}{2}} \sum_{j \in I_u}y_j\right) $
# - $ \pmb{I_u}$ --- the set of all items rated by user u
#
# - $\pmb{y_j}$ --- Our new set of item factors that capture implicit ratings.
# - __Optimization problem with user item interactions and regularization (to avoid overfitting)__
# -
# - $ \large \sum_{r_{ui} \in R_{train}} \left(r_{ui} - \hat{r}_{ui} \right)^2 +
# \lambda\left(b_i^2 + b_u^2 + ||q_i||^2 + ||p_u||^2 + ||y_j||^2\right) $
# In[ ]:
# initiallize the model
svdpp = SVDpp(n_factors=50, random_state=15, verbose=True)
svdpp_train_results, svdpp_test_results = run_surprise(svdpp, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
models_evaluation_train['svdpp'] = svdpp_train_results
models_evaluation_test['svdpp'] = svdpp_test_results
#
#
# 4.4.7 XgBoost with 13 features + Surprise Baseline + Surprise KNNbaseline + MF Techniques
# __Preparing Train data__
# In[ ]:
# add the predicted values from both knns to this dataframe
reg_train['svd'] = models_evaluation_train['svd']['predictions']
reg_train['svdpp'] = models_evaluation_train['svdpp']['predictions']
reg_train.head(2)
# __Preparing Test data __
# In[ ]:
reg_test_df['svd'] = models_evaluation_test['svd']['predictions']
reg_test_df['svdpp'] = models_evaluation_test['svdpp']['predictions']
reg_test_df.head(2)
#
# In[ ]:
# prepare x_train and y_train
x_train = reg_train.drop(['user', 'movie', 'rating',], axis=1)
y_train = reg_train['rating']
# prepare test data
x_test = reg_test_df.drop(['user', 'movie', 'rating'], axis=1)
y_test = reg_test_df['rating']
xgb_final = xgb.XGBRegressor(n_jobs=10, random_state=15)
train_results, test_results = run_xgboost(xgb_final, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
models_evaluation_train['xgb_final'] = train_results
models_evaluation_test['xgb_final'] = test_results
xgb.plot_importance(xgb_final)
plt.show()
# 4.4.8 XgBoost with Surprise Baseline + Surprise KNNbaseline + MF Techniques
# In[ ]:
# prepare train data
x_train = reg_train[['knn_bsl_u', 'knn_bsl_m', 'svd', 'svdpp']]
y_train = reg_train['rating']
# test data
x_test = reg_test_df[['knn_bsl_u', 'knn_bsl_m', 'svd', 'svdpp']]
y_test = reg_test_df['rating']
xgb_all_models = xgb.XGBRegressor(n_jobs=10, random_state=15)
train_results, test_results = run_xgboost(xgb_all_models, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
models_evaluation_train['xgb_all_models'] = train_results
models_evaluation_test['xgb_all_models'] = test_results
xgb.plot_importance(xgb_all_models)
plt.show()
#
#
# 4.5 Comparision between all models
# In[ ]:
# Saving our TEST_RESULTS into a dataframe so that you don't have to run it again
pd.DataFrame(models_evaluation_test).to_csv('sample/small/small_sample_results.csv')
models = pd.read_csv('sample/small/small_sample_results.csv', index_col=0)
models.loc['rmse'].sort_values()
#
# In[ ]:
print("-"*100)
print("Total time taken to run this entire notebook ( with saved files) is :",datetime.now()-globalstart)
# 5. Assignment
# 1.Instead of using 10K users and 1K movies to train the above models, use 25K users and 3K movies (or more) to train all of the above models. Report the RMSE and MAPE on the test data using larger amount of data and provide a comparison between various models as shown above.
#
# NOTE: Please be patient as some of the code snippets make take many hours to compelte execution.
#
# 2.Tune hyperparamters of all the Xgboost models above to improve the RMSE.
# In[ ]:
get_ipython().run_cell_magic('javascript', '', '// Converts integer to roman numeral\n// https://github.com/kmahelona/ipython_notebook_goodies\n// https://kmahelona.github.io/ipython_notebook_goodies/ipython_notebook_toc.js\nfunction romanize(num) {\n var lookup = {M:1000,CM:900,D:500,CD:400,C:100,XC:90,L:50,XL:40,X:10,IX:9,V:5,IV:4,I:1},\n\troman = \'\',\n\t i;\n\tfor ( i in lookup ) {\n\t while ( num >= lookup[i] ) {\n\t\troman += i;\n\t\tnum -= lookup[i];\n\t }\n\t}\n\treturn roman;\n }\n\n// Builds a Table of Contents from all in DOM\nfunction createTOC(){\n var toc = "";\n var level = 0;\n var levels = {}\n $(\'#toc\').html(\'\');\n\n $(":header").each(function(i){\n\t if (this.id==\'tocheading\'){return;}\n \n\t var titleText = this.innerHTML;\n\t var openLevel = this.tagName[1];\n\n\t if (levels[openLevel]){\n\t\tlevels[openLevel] += 1;\n\t } else{\n\t\tlevels[openLevel] = 1;\n\t }\n\n\t if (openLevel > level) {\n\t\ttoc += (new Array(openLevel - level + 1)).join(\'\');\n\t } else if (openLevel < level) {\n\t\ttoc += (new Array(level - openLevel + 1)).join("
");\n\t\tfor (i=level;i>openLevel;i--){levels[i]=0;}\n\t }\n\n\t level = parseInt(openLevel);\n\n\n\t if (this.id==\'\'){this.id = this.innerHTML.replace(/ /g,"-")}\n\t var anchor = this.id;\n \n\t toc += \'- \' + titleText + \'
\';\n \n\t});\n\n \n if (level) {\n\ttoc += (new Array(level + 1)).join("
");\n }\n\n \n $(\'#toc\').append(toc);\n\n};\n\n// Executes the createToc function\nsetTimeout(function(){createTOC();},100);\n\n// Rebuild to TOC every minute\nsetInterval(function(){createTOC();},60000);\n')
# In[ ]:
tuned_models_evaluation_train = dict()
tuned_models_evaluation_test = dict()
tuned_models_evaluation_train, tuned_models_evaluation_test
#
# 5.1 XGBoost with hyperparameter tuning with initial 13 features
# In[ ]:
param_grid = dict(
learning_rate = [0.0001,0.001,0.01,0.1],
max_depth = [3,4,5,6],
n_estimators = [100,150,200,250]
)
# In[ ]:
from sklearn.model_selection import GridSearchCV
from xgboost.sklearn import XGBRegressor
# ### 5.1 Tuning XGBoost with initial 13 features
# In[ ]:
# prepare Train data
x_train = reg_train.drop(['user','movie','rating'], axis=1)
y_train = reg_train['rating']
# Prepare Test data
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# In[ ]:
# making model ready for tuning
xgb_model = XGBRegressor(silent=False, n_jobs=-1,random_state=15,verbosity=1,nthread=-1)
xgb_knn_bsl = GridSearchCV(xgb_model,param_grid,n_jobs=-1,cv = 3,verbose=10,return_train_score=True)
# In[ ]:
train_results, test_results = run_xgboost(xgb_knn_bsl, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
tuned_models_evaluation_train['first_algo'] = train_results
tuned_models_evaluation_test['first_algo'] = test_results
# Done. Time taken : 0:34:43.292529
#
# Done
#
# Evaluating the model with TRAIN data...
# Evaluating Test data
#
# > TEST DATA Scores
# ------------------------------
# * RMSE : 1.074079123678349
# * MAPE : 34.74543952747899
#
#
#
#
#
# In[ ]:
print(x_train.columns)
print(xgb_knn_bsl.best_estimator_.feature_importances_)
# In[ ]:
import xgboost as xgb
xgb.plot_importance(xgb_knn_bsl.best_estimator_)
plt.show()
# In[ ]:
# ### 5.2 Tuning XGBoost with initial 13 features + Surprise Baseline predictor
# In[ ]:
from surprise import BaselineOnly
# __Predicted_rating : ( baseline prediction ) __
#
# - http://surprise.readthedocs.io/en/stable/basic_algorithms.html#surprise.prediction_algorithms.baseline_only.BaselineOnly
# >$ \large {\hat{r}_{ui} = b_{ui} =\mu + b_u + b_i} $
#
#
# - $\pmb \mu $ : Average of all trainings in training data.
# - $\pmb b_u$ : User bias
# - $\pmb b_i$ : Item bias (movie biases)
# __Optimization function ( Least Squares Problem ) __
#
# - http://surprise.readthedocs.io/en/stable/prediction_algorithms.html#baselines-estimates-configuration
#
# > $ \large \sum_{r_{ui} \in R_{train}} \left(r_{ui} - (\mu + b_u + b_i)\right)^2 +
# \lambda \left(b_u^2 + b_i^2 \right).\text { [mimimize } {b_u, b_i]}$
# In[ ]:
# create the traindata from the dataframe...
train_data = Dataset.load_from_df(reg_train[['user', 'movie', 'rating']], reader)
# build the trainset from traindata.., It is of dataset format from surprise library..
trainset = train_data.build_full_trainset()
testset = list(zip(reg_test_df.user.values, reg_test_df.movie.values, reg_test_df.rating.values))
testset[:3]
# In[ ]:
# options are to specify.., how to compute those user and item biases
bsl_options = {'method': 'sgd',
'learning_rate': .001
}
bsl_algo = BaselineOnly(bsl_options=bsl_options)
# run this algorithm.., It will return the train and test results..
bsl_train_results, bsl_test_results = run_surprise(bsl_algo, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
tuned_models_evaluation_train['bsl_algo'] = bsl_train_results
tuned_models_evaluation_test['bsl_algo'] = bsl_test_results
# __Updating Train Data__
# In[ ]:
# add our baseline_predicted value as our feature..
reg_train['bslpr'] = tuned_models_evaluation_train['bsl_algo']['predictions']
reg_train.head(2)
# __Updating Train Data__
# In[ ]:
# add that baseline predicted ratings with Surprise to the test data as well
reg_test_df['bslpr'] = tuned_models_evaluation_test['bsl_algo']['predictions']
reg_test_df.head(2)
# In[ ]:
#here
# In[ ]:
# prepare train data
x_train = reg_train.drop(['user', 'movie','rating'], axis=1)
y_train = reg_train['rating']
# Prepare Test data
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# making model ready for tuning
xgb_model = XGBRegressor(silent=False, n_jobs=-1,random_state=15,verbosity=1,nthread=-1)
xgb_bsl = GridSearchCV(xgb_model,param_grid,n_jobs=-1,cv = 3,verbose=10,return_train_score=True)
# In[ ]:
train_results, test_results = run_xgboost(xgb_bsl, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
tuned_models_evaluation_train['xgb_bsl'] = train_results
tuned_models_evaluation_test['xgb_bsl'] = test_results
# In[ ]:
xgb.plot_importance(xgb_bsl.best_estimator_)
plt.show()
# In[ ]:
# ### 5.3 Tuning XGBoost with initial 13 features + Surprise Baseline predictor + KNNBaseline predictor
# In[ ]:
from surprise import KNNBaseline
# - KNN BASELINE
# - http://surprise.readthedocs.io/en/stable/knn_inspired.html#surprise.prediction_algorithms.knns.KNNBaseline
# - PEARSON_BASELINE SIMILARITY
# - http://surprise.readthedocs.io/en/stable/similarities.html#surprise.similarities.pearson_baseline
# - SHRINKAGE
# - _2.2 Neighborhood Models_ in http://courses.ischool.berkeley.edu/i290-dm/s11/SECURE/a1-koren.pdf
# - __predicted Rating__ : ( ___ based on User-User similarity ___ )
#
# \begin{align} \hat{r}_{ui} = b_{ui} + \frac{ \sum\limits_{v \in N^k_i(u)}
# \text{sim}(u, v) \cdot (r_{vi} - b_{vi})} {\sum\limits_{v \in
# N^k_i(u)} \text{sim}(u, v)} \end{align}
#
# - $\pmb{b_{ui}}$ - _Baseline prediction_ of (user,movie) rating
#
# - $ \pmb {N_i^k (u)}$ - Set of __K similar__ users (neighbours) of __user (u)__ who rated __movie(i)__
#
# - _sim (u, v)_ - __Similarity__ between users __u and v__
# - Generally, it will be cosine similarity or Pearson correlation coefficient.
# - But we use __shrunk Pearson-baseline correlation coefficient__, which is based on the pearsonBaseline similarity ( we take base line predictions instead of mean rating of user/item)
#
#
# - __ Predicted rating __ ( based on Item Item similarity ):
# \begin{align} \hat{r}_{ui} = b_{ui} + \frac{ \sum\limits_{j \in N^k_u(i)}\text{sim}(i, j) \cdot (r_{uj} - b_{uj})} {\sum\limits_{j \in N^k_u(j)} \text{sim}(i, j)} \end{align}
#
# - ___Notations follows same as above (user user based predicted rating ) ___
# - Surprise KNNBaseline with user user similarities
# 4.4.4.1 Surprise KNNBaseline with user user similarities
# In[ ]:
# we specify , how to compute similarities and what to consider with sim_options to our algorithm
sim_options = {'user_based' : True,
'name': 'pearson_baseline',
'shrinkage': 100,
'min_support': 2
}
# we keep other parameters like regularization parameter and learning_rate as default values.
bsl_options = {'method': 'sgd'}
knn_bsl_u = KNNBaseline(k=40, sim_options = sim_options, bsl_options = bsl_options)
knn_bsl_u_train_results, knn_bsl_u_test_results = run_surprise(knn_bsl_u, trainset, testset, verbose=True)
# In[ ]:
# Just store these error metrics in our models_evaluation datastructure
tuned_models_evaluation_train['knn_bsl_u'] = knn_bsl_u_train_results
tuned_models_evaluation_test['knn_bsl_u'] = knn_bsl_u_test_results
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","wb") as f:
pickle.dump(tuned_models_evaluation_train,f)
pickle.dump(tuned_models_evaluation_test,f)
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","rb") as f:
tuned_models_evaluation_train = pickle.load(f)
tuned_models_evaluation_test = pickle.load(f)
# - Surprise KNNBaseline with movie movie similarities
# In[ ]:
# we specify , how to compute similarities and what to consider with sim_options to our algorithm
# 'user_based' : Fals => this considers the similarities of movies instead of users
sim_options = {'user_based' : False,
'name': 'pearson_baseline',
'shrinkage': 100,
'min_support': 2
}
# we keep other parameters like regularization parameter and learning_rate as default values.
bsl_options = {'method': 'sgd'}
knn_bsl_m = KNNBaseline(k=40, sim_options = sim_options, bsl_options = bsl_options)
knn_bsl_m_train_results, knn_bsl_m_test_results = run_surprise(knn_bsl_m, trainset, testset, verbose=True)
# In[ ]:
# Just store these error metrics in our models_evaluation datastructure
tuned_models_evaluation_train['knn_bsl_m'] = knn_bsl_m_train_results
tuned_models_evaluation_test['knn_bsl_m'] = knn_bsl_m_test_results
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","wb") as f:
pickle.dump(tuned_models_evaluation_train,f)
pickle.dump(tuned_models_evaluation_test,f)
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","rb") as f:
tuned_models_evaluation_train = pickle.load(f)
tuned_models_evaluation_test = pickle.load(f)
# In[ ]:
# add the predicted values from both knns to this dataframe
reg_train['knn_bsl_u'] = tuned_models_evaluation_train['knn_bsl_u']['predictions']
reg_train['knn_bsl_m'] = tuned_models_evaluation_train['knn_bsl_m']['predictions']
reg_train.head(2)
# In[ ]:
reg_test_df.head(2)
# In[ ]:
reg_test_df['knn_bsl_u'] = tuned_models_evaluation_test['knn_bsl_u']['predictions']
reg_test_df['knn_bsl_m'] = tuned_models_evaluation_test['knn_bsl_m']['predictions']
reg_test_df.head(2)
# In[ ]:
# prepare the train data....
x_train = reg_train.drop(['user', 'movie', 'rating'], axis=1)
y_train = reg_train['rating']
# prepare the train data....
x_test = reg_test_df.drop(['user','movie','rating'], axis=1)
y_test = reg_test_df['rating']
# making model ready for tuning
xgb_model = XGBRegressor(silent=False, n_jobs=-1,random_state=15,verbosity=1,nthread=-1)
xgb_knn_bsl_um = GridSearchCV(xgb_model,param_grid,n_jobs=-1,cv = 3,verbose=10,return_train_score=True)
# In[ ]:
train_results, test_results = run_xgboost(xgb_knn_bsl_um, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
tuned_models_evaluation_train['xgb_knn_bsl_um'] = train_results
tuned_models_evaluation_test['xgb_knn_bsl_um'] = test_results
# Done. Time taken : 0:53:27.148315
#
# Done
# In[ ]:
y_train_pred = xgb_knn_bsl_um.predict(x_train)
# get the rmse and mape of train data...
rmse_train, mape_train = get_error_metrics(y_train.values, y_train_pred)
# store the results in train_results dictionary..
train_results = {'rmse': rmse_train,
'mape' : mape_train,
'predictions' : y_train_pred}
print('\nTrain DATA')
print('-'*30)
print('RMSE : ', rmse_train)
print('MAPE : ', mape_train)
# In[ ]:
print('Evaluating Test data')
y_test_pred = xgb_knn_bsl_um.predict(x_test)
rmse_test, mape_test = get_error_metrics(y_true=y_test.values, y_pred=y_test_pred)
# store them in our test results dictionary.
test_results = {'rmse': rmse_test,
'mape' : mape_test,
'predictions':y_test_pred}
if verbose:
print('\nTEST DATA')
print('-'*30)
print('RMSE : ', rmse_test)
print('MAPE : ', mape_test)
# In[ ]:
tuned_models_evaluation_train['xgb_knn_bsl_um'] = train_results
tuned_models_evaluation_test['xgb_knn_bsl_um'] = test_results
# In[ ]:
xgb.plot_importance(xgb_knn_bsl_um.best_estimator_)
# In[ ]:
from surprise import SVD
# ### 5.4 XgBoost with Surprise Baseline + Surprise KNNbaseline + MF Techniques
# - svd
# In[ ]:
# initiallize the model
svd = SVD(n_factors=100, biased=True, random_state=15, verbose=True)
svd_train_results, svd_test_results = run_surprise(svd, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
tuned_models_evaluation_train['svd'] = svd_train_results
tuned_models_evaluation_test['svd'] = svd_test_results
# - svd++
# In[ ]:
from surprise import SVDpp
# In[ ]:
# initiallize the model
svdpp = SVDpp(n_factors=50, random_state=15, verbose=True)
svdpp_train_results, svdpp_test_results = run_surprise(svdpp, trainset, testset, verbose=True)
# Just store these error metrics in our models_evaluation datastructure
tuned_models_evaluation_train['svdpp'] = svdpp_train_results
tuned_models_evaluation_test['svdpp'] = svdpp_test_results
# In[ ]:
# add the predicted values from both knns to this dataframe
reg_train['svd'] =tuned_models_evaluation_train['svd']['predictions']
reg_train['svdpp'] = tuned_models_evaluation_train['svdpp']['predictions']
reg_train.head(2)
# In[ ]:
reg_train.shape
# In[ ]:
reg_test_df['svd'] = tuned_models_evaluation_test['svd']['predictions']
reg_test_df['svdpp'] = tuned_models_evaluation_test['svdpp']['predictions']
reg_test_df.head(2)
# In[ ]:
# prepare x_train and y_train
x_train = reg_train.drop(['user', 'movie', 'rating',], axis=1)
y_train = reg_train['rating']
# prepare test data
x_test = reg_test_df.drop(['user', 'movie', 'rating'], axis=1)
y_test = reg_test_df['rating']
# making model ready for tuning
xgb_model = XGBRegressor(silent=False, n_jobs=-1,random_state=15,verbosity=1,nthread=-1)
xgb_final = GridSearchCV(xgb_model,param_grid,n_jobs=-1,cv = 3,verbose=10,return_train_score=True)
# In[ ]:
train_results, test_results = run_xgboost(xgb_final, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
tuned_models_evaluation_train['xgb_final'] = train_results
tuned_models_evaluation_test['xgb_final'] = test_results
# In[ ]:
xgb.plot_importance(xgb_final.best_estimator_)
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","wb") as f:
pickle.dump(tuned_models_evaluation_train,f)
pickle.dump(tuned_models_evaluation_test,f)
# ### 5.5 XgBoost with Surprise Baseline + Surprise KNNbaseline + MF Techniques
# In[ ]:
# prepare train data
x_train = reg_train[['knn_bsl_u', 'knn_bsl_m', 'svd', 'svdpp']]
y_train = reg_train['rating']
# test data
x_test = reg_test_df[['knn_bsl_u', 'knn_bsl_m', 'svd', 'svdpp']]
y_test = reg_test_df['rating']
# making model ready for tuning
xgb_model = XGBRegressor(silent=False, n_jobs=-1,random_state=15,verbosity=1,nthread=-1)
xgb_all_models = GridSearchCV(xgb_model,param_grid,n_jobs=-1,cv = 3,verbose=10,return_train_score=True)
# In[ ]:
train_results, test_results = run_xgboost(xgb_all_models, x_train, y_train, x_test, y_test)
# store the results in models_evaluations dictionaries
tuned_models_evaluation_train['xgb_all_models'] = train_results
tuned_models_evaluation_test['xgb_all_models'] = test_results
# In[ ]:
xgb.plot_importance(xgb_all_models.best_estimator_)
# In[ ]:
import pickle
with open("/content/drive/My Drive/Netflix_recommender/tuned_result.pkl","wb") as f:
pickle.dump(tuned_models_evaluation_train,f)
pickle.dump(tuned_models_evaluation_test,f)
# In[ ]:
# Saving our TEST_RESULTS into a dataframe so that you don't have to run it again
pd.DataFrame(tuned_models_evaluation_test).to_csv('drive/My Drive/Netflix_recommender/tuned_small_sample_results.csv')
models = pd.read_csv('drive/My Drive/Netflix_recommender/tuned_small_sample_results.csv', index_col=0)
models.loc['rmse'].sort_values()
# In[ ]:
df_25k_train = pd.DataFrame(data=tuned_models_evaluation_train)
df_25k_train.drop("predictions",inplace = True)
df_25k_train
# In[ ]:
df_25k_test = pd.DataFrame(data=tuned_models_evaluation_test)
df_25k_test.drop("predictions",inplace = True)
df_25k_test
# ## **Note:**
# ### All the results are computed on 25k data points
# ## Observations
#
# 1. We have got slight better scores as compared to scores on 10k data.
# 2. Compute time has increased drasticaly as we have increased points.
# 3. We have plot feature_importance for each models that helps in understanding features.
# 4. **knn_bsl_u and knn_bsl_m** have got slightly better scores as compared to other models.
# 5. Surprise library features yeild similar as compared with other Self computed features.