app.py

#!/usr/bin/env python
# coding: utf-8

# ## Data Loading

# Importing the necessary libraries, like pandas, numpy and some plotting libraries such as matplotlib and seaborn

# In[ ]:


import pandas as pd
import numpy as np 
import matplotlib
import matplotlib.pyplot as plt
import seaborn as sns
get_ipython().run_line_magic('matplotlib', 'inline')


# Set the default font size, figure size and the grid in the plot

# In[ ]:


sns.set_style('darkgrid')
matplotlib.rcParams['font.size'] = 14
matplotlib.rcParams['figure.figsize'] = (10, 6)
matplotlib.rcParams['figure.facecolor'] = '#00000000'


# Reading of data as a pandas dataframe and named as **df**

# In[ ]:


df = pd.read_csv('Walmart.csv')


# In[ ]:


df


# **About Data:**
# *   Store - the store number
# *   Date - the week of sales
# *   Weekly_Sales - sales for the given store
# *   Holiday_Flag - whether the week is a special holiday week 1 – Holiday week 0 – Non-holiday week
# *   Temperature - Temperature on the day of sale
# *   Fuel_Price - Cost of fuel in the region
# *   CPI – Prevailing consumer price index
# *   Unemployment - Prevailing unemployment rate

# **Insights:**
# 
# *   Here the target columns is Weekly_Sales.
# *   The data is related to walmart store of united state of america. Where **Store**, **Holiday_Flag** are categorical in nature
# *   The data is collected over a 45 stores and weekly sales gives the sales of the crossponding store.
# 

# ## Data Exploration and Modification

# In[ ]:


df.info() # it gives the information (like count and data type) of the dataset


# Here Date columns is **object** and other remain columns are **interger or float** in nature. Now using the pandas I change the date column datatype(i.e. object) into a pandas-datetime. 

# In[ ]:


df.Date=pd.to_datetime(df.Date)


# Using the date column i create three seperate columns of weekday, month and year and added to the existing dataset.

# In[ ]:


df['weekday'] = df.Date.dt.weekday
df['month'] = df.Date.dt.month
df['year'] = df.Date.dt.year


# Now I drop the date columns because of no use of it.

# In[ ]:


df.drop(['Date'], axis=1, inplace=True)


# Hence the modified dataset is look like:

# In[ ]:


df.head(3)


# Explored the unique values of the weekday, month and year columns as follows:

# In[ ]:


print('years unique value', df.year.unique())
print('months unique value', df.month.unique())
print('weekday unique value', df.weekday.unique())


# Months and weekday are as usual, but the data is taken from year 2010, 2011, 2012 only.

# Now to get the idea of distribution of the dataset, I used describe function which gives a table of various statistical values of all the columns

# In[ ]:


df.describe()


# **Insights:**
# *   Temperature - has values ranges from (-2, 100.1) Fahrenhite.
# *   CPI - is ranges from 126 to 227 with a standard deviation of 39.35
# *   Unemployment - is ranges from 3.87 to 14.31 with a standard deviation of 1.87

# In[ ]:


original_df = df.copy() # made the copy of dataframe to check the dublicates values in the dataset


# Checking of dublicates values : 

# In[ ]:


counter = 0
rs,cs = original_df.shape

df.drop_duplicates(inplace=True)

if df.shape==(rs,cs):
    print('The dataset doesn\'t have any duplicates')
else:
    print('Number of duplicates dropped/fixed ---> {rs-df.shape[0]}')


# Checking of missing values : 

# In[ ]:


df.isnull().sum()


# Dataset doesn't have null values

# ## Data Visualization

# In[ ]:


df.head(3)


# Here we have:
# 
# **Numerical columns:** Weekly_sales, temperature, fuel_price, cpi, unemployment
# 
# **Categorical columns:** Holiday_flag, Weekday, month, year
# 
# Now plotted the count plot to get the distribution or frequency of the columns

# In[ ]:


fig, axes = plt.subplots(2, 2, figsize=(16, 8))

#axes[0,0].set_title('Holiday Count plot')
sns.countplot(x='Holiday_Flag', data=df, ax= axes[0,0])

#axes[0,1].set_title('Weekday Count plot')
sns.countplot(x='weekday', data=df, ax= axes[0,1]);

#axes[1,0].set_title('month Count plot')
sns.countplot(x='month', data=df, ax= axes[1,0]);

#axes[1,1].set_title('year Count plot')
sns.countplot(x='year', data=df, ax= axes[1,1]);


# **Insights:**
# 
# *   In Holiday flag most of the time there is no holiday in that week.
# *   In weekdays columns observations are mostly related to the day 4
# *   Most of the observation in the data is from the month of april
# *   Most of the observation in the data is from year 2011

# To get the idea of how many observations are there in dataset crossponding to each store, I again plot a count plot.

# In[ ]:


plt.figure(figsize= (18,8))
sns.countplot(x= 'Store', data= df);
plt.show()


# All the store have equal number of data in the set 

# In[ ]:


df.head(1)


# To analyze the distribution of the data, I plotted the histogram and boxplot for Temperature, Unemployment, Fuel_Price, CPI.

# In[ ]:


fig, axes = plt.subplots(4, 2, figsize=(16, 16))
# axes[0,0].set_title('Temperature')
sns.histplot(x= 'Temperature', data= df, ax= axes[0,0])

sns.boxplot(x= 'Temperature', data= df, ax= axes[0,1])

# axes[1,0].set_title('Unemployment')
sns.histplot(x= 'Unemployment', data= df, ax= axes[1,0])

sns.boxplot(x= 'Unemployment', data= df, ax= axes[1,1])

# axes[2,0].set_title('Fuel_Price')
sns.histplot(x= 'Fuel_Price', data= df, ax= axes[2,0])

sns.boxplot(x = 'Fuel_Price', data= df, ax= axes[2,1])

# axes[3,0].set_title('CPI')
sns.histplot(x= 'CPI', data= df, ax= axes[3,0])

sns.boxplot(x= 'CPI', data= df, ax= axes[3,1]);


# **Insights:**
# 
# *   Temperature:  Crossponding to the lower temperature, there is a presence of outlier.
# *   Umemployment: The outlier is present in the dataset crossponding to higher and lower both values.
# *   CPI: It is either very low or very high.

# In[ ]:


# Removing the outlier from Temperature column

Q1 = df['Temperature'].quantile(0.25)
Q3 = df['Temperature'].quantile(0.75)
IQR = Q3 - Q1
df = df[df['Temperature'] <= (Q3+(1.5*IQR))]
df = df[df['Temperature'] >= (Q1-(1.5*IQR))]


# In[ ]:


# Removing the outlier from Unemployment column

Q1 = df['Unemployment'].quantile(0.25)
Q3 = df['Unemployment'].quantile(0.75)
IQR = Q3 - Q1
df = df[df['Unemployment'] <= (Q3+(1.5*IQR))]
df = df[df['Unemployment'] >= (Q1-(1.5*IQR))]


# In[ ]:


df.shape


# On the process of removing outlier, **484 data** points are removed from data-set

# ## Encoding

# Encoding is a process to convert the categorical columns into a numerical columns, as it is not a good preactice to train a model with categorical inputs.

# In[ ]:


cat_cols = ['Store', 'Holiday_Flag', 'weekday', 'month', 'year'] # these are the categorical columns


# In[ ]:


df[cat_cols].nunique() # Counting the unique value in each of the categorical columns.


# In[ ]:


# Imported OneHotEncoder to perfrom the encoding
from sklearn.preprocessing import OneHotEncoder
# Creating a object of the encoder function
encoder = OneHotEncoder(sparse=False, handle_unknown='ignore')
# Fit the encoder object to the dataset which i want to convert into numerical form. 
encoder.fit(df[cat_cols])


# In[ ]:


# Creating a list of the encoded columns 
encoded_cols = list(encoder.get_feature_names(cat_cols))
print(encoded_cols)


# In[ ]:


# Now i added those encoded columns into the original dataset by transforming it into a categorical form.
df[encoded_cols] = encoder.transform(df[cat_cols])


# In[ ]:


df.shape


# ## Standardization

# To scale all the column values to specific range of 0 - 1, I used standard scaler function. It is important to give the equal weights to all the columns.

# In[ ]:


# Importing a MinMaxScaler
from sklearn.preprocessing import MinMaxScaler
# Creating Scaler Object
scaler = MinMaxScaler()
# Fitted the scaler to the dataset
scaler.fit(df)
# Transformed the dataset using the fitted scaler object 
scaled_df = scaler.transform(df)


# In[ ]:


# Converting the output scaled dataframe into a pandas dataframe
scaled_df = pd.DataFrame(data = scaled_df, columns = df.columns)


# In[ ]:


# Checking the output dataframe
scaled_df.head(3)


# ## Train-Test-Split

# Split the dataset into the two part: 
# 1. Training dataset (used to train the model)
# 2. Testing dataset (used to test the model)

# In[ ]:


# Drop the sales columns to get the input  features
X = scaled_df.drop('Weekly_Sales', axis=1)
# Use the sales column as a target columns
y = scaled_df['Weekly_Sales']


# In[ ]:


# Importing train test split
from sklearn.model_selection import train_test_split
# dividing the dataset into the train and the test parts and each part has input feature and target features
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state = 42)


# In[ ]:


# Printin the shape of all the dataset
X_train.shape, X_test.shape, y_train.shape, y_test.shape


# ## Feature Selection

# Out of all the 78 features all are not important and we have to choose the important feature out of all the features

# In[ ]:


# import a linear regerssion model
from sklearn.linear_model import LinearRegression
# import a Random Forest Regressor model
from sklearn.ensemble import RandomForestRegressor
# import a mean squared error for model evaluation
from sklearn.metrics import mean_squared_error
# import a r2 score for model evaluation
from sklearn.metrics import r2_score
# import a RFE model for feature selection
from sklearn.feature_selection import RFE


# In[ ]:


# Creatint a list to store training and test error 
Trr=[]; Tss=[]; n=3
order=['ord-'+str(i) for i in range(2,n)]
Trd = pd.DataFrame(np.zeros((10,n-2)), columns=order)
Tsd = pd.DataFrame(np.zeros((10,n-2)), columns=order)

m=df.shape[1]-2
for i in range(m):
    # creating a linear regression model object
    lm = LinearRegression()
    # creating a rfe model object with linear regression model and with a parameter of the number of features 
    rfe = RFE(lm, n_features_to_select=X_train.shape[1]-i)
    # fitting the rfe model to the trainig dataset
    rfe = rfe.fit(X_train, y_train)
    # creating a linear regression model object for prediction
    LR = LinearRegression()
    # fitted the lr model using the selected features
    LR.fit(X_train.loc[:,rfe.support_], y_train)
    # Made the prediction using the linear regression model
    pred1 = LR.predict(X_train.loc[:,rfe.support_]) # make the prediction on the trainig dataset
    pred2 = LR.predict(X_test.loc[:,rfe.support_]) # make the prediction on the test dataset
    # Insert the mse into the Trr and Tss for train and test respectively
    Trr.append(np.sqrt(mean_squared_error(y_train, pred1)))
    Tss.append(np.sqrt(mean_squared_error(y_test, pred2)))


# In[ ]:


plt.plot(Trr, label= 'Train RMSE')
plt.plot(Tss, label= 'Test RMSE')
plt.legend()
plt.show()


# If we Recursively Eleminate at most **Ten** features then the score is maximum.

# In[2]:


# Eleminating 10 features and using Linear Regresion model the error printed as follows which is the best possible score.

# creating a linear regression model object
lm = LinearRegression()
# creating a rfe model object with linear regression model and with number of features equal to 10.
rfe = RFE(lm,n_features_to_select=X_train.shape[1]-9) 
# fitting the rfe model to the trainig dataset
rfe = rfe.fit(X_train, y_train)
# creating a linear regression model object for prediction
LR = LinearRegression()
# fitted the lr model using the selected features
LR.fit(X_train.loc[:,rfe.support_], y_train)
# Made the prediction using the linear regression model
pred1 = LR.predict(X_train.loc[:,rfe.support_])
pred2 = LR.predict(X_test.loc[:,rfe.support_])
# Printing the results as a MSE and r2_score.
print("MSE train",np.sqrt(mean_squared_error(y_train, pred1)))
print("MSE test",np.sqrt(mean_squared_error(y_test, pred2)))
print("r2_score train - {}".format(r2_score(y_train, pred1)))
print("r2_score test - {}".format(r2_score(y_test, pred2)))


# Now Removing the 10 features and create the New training and test dataset

# In[ ]:


X_train = X_train.loc[:,rfe.support_]
X_test = X_test.loc[:,rfe.support_]


# Now onwards I am going to use various models 

# ## Linear Regression

# In[ ]:


lr =LinearRegression()
lr.fit(X_train, y_train)
pred1 = lr.predict(X_train)
pred2 = lr.predict(X_test)

print("Root Mean Squared Error train {}".format(np.mean(mean_squared_error(y_train, pred1))))
print("Root Mean Squared Error test {}".format(np.sqrt(mean_squared_error(y_test, pred2))))
print("r2_score train {}".format(r2_score(y_train, pred1)))
print("r2_score test {}".format(r2_score(y_test, pred2)))


# **Ridge Regression**

# In[ ]:


from sklearn.linear_model import Ridge
rr = Ridge()
rr.fit(X_train, y_train)
predrr1 = rr.predict(X_train)
predrr2 = rr.predict(X_test)
print("Root Mean Squared Error train {}".format(np.mean(mean_squared_error(y_train, predrr1))))
print("Root Mean Squared Error test {}".format(np.sqrt(mean_squared_error(y_test, predrr2))))
print("r2_score train {}".format(r2_score(y_train, predrr1)))
print("r2_score test {}".format(r2_score(y_test, predrr2)))


# **Lasso Regression**

# In[ ]:


from sklearn.linear_model import Lasso
lr = Lasso()
lr.fit(X_train, y_train)
predlr1 = lr.predict(X_train)
predlr2 = lr.predict(X_test)
print("Root Mean Squared Error train {}".format(np.mean(mean_squared_error(y_train, predlr1))))
print("Root Mean Squared Error test {}".format(np.sqrt(mean_squared_error(y_test, predlr2))))
print("r2_score train {}".format(r2_score(y_train, predlr1)))
print("r2_score test {}".format(r2_score(y_test, predlr2)))


# **ElasticNet Regression**

# In[ ]:


from sklearn.linear_model import ElasticNet
en = ElasticNet()
en.fit(X_train, y_train)
predlr1 = en.predict(X_train)
predlr2 = en.predict(X_test)
print("Root Mean Squared Error train {}".format(np.mean(mean_squared_error(y_train, predlr1))))
print("Root Mean Squared Error test {}".format(np.sqrt(mean_squared_error(y_test, predlr2))))
print("r2_score train {}".format(r2_score(y_train, predlr1)))
print("r2_score test {}".format(r2_score(y_test, predlr2)))


# **Polynomial Regression**

# In[ ]:


from sklearn.preprocessing import PolynomialFeatures


# In[ ]:


Trr = []
Tss = []
for i in range(2,4):
  poly_reg = PolynomialFeatures(degree = i)
  pl_X_train = poly_reg.fit_transform(X_train)
  pl_X_test = poly_reg.fit_transform(X_test)
  lr = LinearRegression()
  lr.fit(pl_X_train, y_train)
  pred_poly_train = lr.predict(pl_X_train)
  Trr.append(np.sqrt(mean_squared_error(y_train, pred_poly_train)))
  pred_poly_test = lr.predict(pl_X_test)
  Tss.append(np.sqrt(mean_squared_error(y_test, pred_poly_test)))


# In[ ]:


plt.figure(figsize=[15,6])
plt.subplot(1,2,1)
plt.plot(range(2,4), Trr, label= 'Training')
plt.plot(range(2,4), Tss, label= 'Testing')
plt.title('Polynomial Feature on training data')
plt.xlabel('Degree')
plt.ylabel('RMSE')
plt.legend()


# It is clear that in between 2-4 degree polynomial regression 2 has Bais-variance tradeoff

# In[ ]:


poly_reg = PolynomialFeatures(degree = 2)
pl_X_train = poly_reg.fit_transform(X_train)
pl_X_test = poly_reg.fit_transform(X_test)
lr = LinearRegression()
lr.fit(pl_X_train, y_train)
pred_poly_train = lr.predict(pl_X_train)
print("r2_score train {}".format(r2_score(pred_poly_train, y_train)))
pred_poly_test = lr.predict(pl_X_test)
print("r2_score test {}".format(r2_score(pred_poly_test, y_test)))
print("Root Mean Squared Error train {}".format(np.mean(mean_squared_error(y_train, pred_poly_train))))
print("Root Mean Squared Error test {}".format(np.sqrt(mean_squared_error(y_test, pred_poly_test))))


# In[ ]:


#creating a tabel 
tabel = {
        'Train R2': [0.9324387485162124, 0.9323641360074176, 0.0, 0.0, 0.9563932198334125],
        'Test R2' : [0.9223162582948724, 0.9219331606995953, -0.00014816618161050954, -0.00014816618161050954, -0.0005599911350040454],
        'Train RMSE' : [0.0016695395619648289, 0.0016713833486400986, 0.024711495499242828, 0.024711495499242828, 0.0010346077251656776 ],
        'Test RMSE' : [0.04569350618906344, 0.04580603645234492, 0.16395383804559885, 0.16395383804559885, 730742413.004261 ]
        }


# In[ ]:


df_new = pd.DataFrame(tabel) 


# In[ ]:


df_new.index = ['Linear Regression', 'Ridge Regression', 'Lasso Regression', 'ElasticNet Regression', 'Polynomial Regression']


# In[ ]:


df_new


# It is clear that Linear Regression is the Best Model in the dataset, with test accuracy of 92%(approx). 
# 
# To improve the accuracy further we can apply other regressor i.e. Random Forest, G

# Now I am going to imporve the accuracy till 98% - 99%. For this I have to use Decision Tree or Random Forest etc.

# **Decision Tree Regressor**

# In[ ]:


from sklearn.tree import DecisionTreeRegressor
dt = DecisionTreeRegressor()
dt.fit(X_train, y_train)


# In[ ]:


pred_dt1 = dt.predict(X_train)
pred_dt2 = dt.predict(X_test)
print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_dt1))))
print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_dt2))))
print('Accuracy Score train: ', dt.score(X_train, y_train))
print('Accuracy Score test: ', dt.score(X_test, y_test))


# In[ ]:


max_depth_range = np.arange(1,40,1)
for x in max_depth_range:
  dt = DecisionTreeRegressor(max_depth= x)
  dt.fit(X_train, y_train)
  pred_dt1 = dt.predict(X_train)
  pred_dt2 = dt.predict(X_test)
  print('for max_depth: ', x)
  print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_dt1))))
  print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_dt2))))
  print('Accuracy Score train: ', dt.score(X_train, y_train))
  print('Accuracy Score test: ', dt.score(X_test, y_test))
  print()


# Decision Tree has maximum accuracy at **maximum depth 39** 

# **Random Forest Regressor**

# In[ ]:


from sklearn.ensemble import RandomForestRegressor
rfc = RandomForestRegressor()
rfc.fit(X_train, y_train)


# In[ ]:


pred_rfc1 = rfc.predict(X_train)
pred_rfc2 = rfc.predict(X_test)
print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_rfc1))))
print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_rfc2))))
print('Accuracy Score train: ', dt.score(X_train, y_train))
print('Accuracy Score test: ', dt.score(X_test, y_test))


# In[ ]:


max_depth_range = np.arange(1,40,1)
for x in max_depth_range:
  dt = RandomForestRegressor(max_depth= x)
  dt.fit(X_train, y_train)
  pred_xg1 = dt.predict(X_train)
  pred_xg2 = dt.predict(X_test)
  print('for max_depth: ', x)
  print('for max_depth: ', x)
  print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_xg1))))
  print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_xg2))))
  print('Accuracy Score train: ', dt.score(X_train, y_train))
  print('Accuracy Score test: ', dt.score(X_test, y_test))
  print()


# In the depth of **36** the** Random Forest Regressor** has its maximum value of accuracy.

# In[ ]:


rfc = RandomForestRegressor(max_depth = 36)
rfc.fit(X_train, y_train)
pred_rfc1 = rfc.predict(X_train)
pred_rfc2 = rfc.predict(X_test)
print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_rfc1))))
print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_rfc2))))
print('Accuracy Score train: ', rfc.score(X_train, y_train))
print('Accuracy Score test: ', rfc.score(X_test, y_test))


# **XG Boost Regressor**

# In[ ]:


from xgboost import XGBRegressor
xg = XGBRegressor()
xg.fit(X_train, y_train)


# In[ ]:


pred_xg1 = xg.predict(X_train)
pred_xg2 = xg.predict(X_test)
print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_xg1))))
print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_xg2))))


# In[ ]:


max_depth_range = np.arange(1,15,1)
for x in max_depth_range:
  dt = XGBRegressor(max_depth= x)
  dt.fit(X_train, y_train)
  pred_xg1 = dt.predict(X_train)
  pred_xg2 = dt.predict(X_test)
  print('for max_depth: ', x)
  print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_xg1))))
  print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_xg2))))
  print('Accuracy Score train: ', dt.score(X_train, y_train))
  print('Accuracy Score test: ', dt.score(X_test, y_test))
  print()


# It means **maximun depth 9** has best value of Accuracy

# In[ ]:


xg = XGBRegressor(max_depth = 9)
xg.fit(X_train, y_train)
pred_xg1 = xg.predict(X_train)
pred_xg2 = xg.predict(X_test)
print("RMSE for train {}".format(np.sqrt(mean_squared_error(y_train, pred_xg1))))
print("RMSE for test {}".format(np.sqrt(mean_squared_error(y_test, pred_xg2))))
print('Accuracy Score train: ', xg.score(X_train, y_train))
print('Accuracy Score test: ', xg.score(X_test, y_test))


# In[ ]:


tabel1 = {
        'Train Score': [0.9679040861170889, 0.9637587048322853, 0.9601543222728802], 
        'Test Score' : [0.8808466556220073, 0.9028060343874318, 0.9115195955339979],
        'Train RMSE' : [0.02816270639447925, 0.02992618589836856, 0.03137907401148098],
        'Test RMSE' :  [0.05659037012899937, 0.051110374979016944, 0.04876553192516943]
        }


# In[ ]:


df1 = pd.DataFrame(tabel1) 


# In[ ]:


df1


# In[ ]:


df1.index = ['Decision Tree', 'Random Forest', 'XGBoost']


# In[ ]:


df1


# Among the method XGBoost is the best method for the data set

# By Comparising the Linear and XGBoost we can conclude that linear Regression the best suited for the above data set

# In[ ]: