Supermarket Regression Notebook

# reference: https://github.com/risenW/medium_tutorial_notebooks/blob/master/supermarket_regression.ipynb
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

# makes graph display in notebook
%matplotlib inline

supermarket_data = pd.read_csv('https://raw.githubusercontent.com/risenW/medium_tutorial_notebooks/master/train.csv')

supermarket_data.head()

supermarket_data.describe()

# remove ID columns
cols_2_remove = ['Product_Identifier', 'Supermarket_Identifier', 'Product_Supermarket_Identifier']

newdata = supermarket_data.drop(cols_2_remove, axis=1)

newdata.head()

cat_cols = ['Product_Fat_Content','Product_Type',
            'Supermarket _Size', 'Supermarket_Location_Type',
           'Supermarket_Type' ]

num_cols = ['Product_Weight', 'Product_Shelf_Visibility',
            'Product_Price', 'Supermarket_Opening_Year', 'Product_Supermarket_Sales']

# bar plot for categorial features
for col in cat_cols:
    fig = plt.figure(figsize=(6,6)) # define plot area
    ax = fig.gca() # define axis  
    
    counts = newdata[col].value_counts() # find the counts for each unique category
    counts.plot.bar(ax = ax) # use the plot.bar method on the counts data frame
    ax.set_title('Bar plot for ' + col)

# scatter plot for numerical features
for col in num_cols:
    fig = plt.figure(figsize=(6,6)) # define plot area
    ax = fig.gca() # define axis  

    newdata.plot.scatter(x = col, y = 'Product_Supermarket_Sales', ax = ax)

# box plot for categorial features
for col in cat_cols:
    sns.boxplot(x=col, y='Product_Supermarket_Sales', data=newdata)
    plt.xlabel(col)
    plt.ylabel('Product Supermarket Sales')
    plt.show() 

# correlation matrix
corrmat = newdata.corr()
f,ax = plt.subplots(figsize=(5,4))
sns.heatmap(corrmat, square=True)

# pair plot of columns without missing values
import warnings
warnings.filterwarnings('ignore')

cat_cols_pair = ['Product_Fat_Content','Product_Type','Supermarket_Location_Type']

cols_2_pair = ['Product_Fat_Content',
             'Product_Shelf_Visibility',
             'Product_Type',
             'Product_Price',
             'Supermarket_Opening_Year',
             'Supermarket_Location_Type',
             'Supermarket_Type',
             'Product_Supermarket_Sales']

for col in cat_cols_pair:
    sns.set()
    plt.figure()
    sns.pairplot(data=newdata[cols_2_pair], height=3.0, hue=col)
    plt.show()

# FEATURE ENGINEERING
# print all unique values
newdata['Product_Fat_Content'].unique()

fat_content_dict = {'Low Fat': 0, 'Ultra Low fat': 0, 'Normal Fat': 1}
newdata['is_normal_fat'] = newdata['Product_Fat_Content'].map(fat_content_dict)

# preview the values
newdata['is_normal_fat'].value_counts()

# assign year 2000 and above as 1, 1996 and below as 0
def cluster_open_year(year):
    if year <= 1996:
        return 0
    else:
        return 1
    
newdata['open_in_the_2000s'] = newdata['Supermarket_Opening_Year'].apply(cluster_open_year)

# preview feature
newdata[['Supermarket_Opening_Year', 'open_in_the_2000s']].head(4)

# get the unique categories in the column as a list
prod_type_cats = list(newdata['Product_Type'].unique())

# remove the class 1 categories
prod_type_cats.remove('Health and Hygiene')
prod_type_cats.remove('Household')
prod_type_cats.remove('Others')

def cluster_prod_type(product):
    if product in prod_type_cats:
        return 0
    else:
        return 1
    
newdata['Product_type_cluster'] = newdata['Product_Type'].apply(cluster_prod_type)

newdata[['Product_Type', 'Product_type_cluster']].tail(10)

# transforming skewed features
fig, ax = plt.subplots(1,2)

# plot of normal Product_Supermarket_Sales on the first axis
sns.histplot(data=newdata['Product_Supermarket_Sales'], bins=15, ax=ax[0])

# transform the Product_Supermarket_Sales and plot on the second axis
newdata['Product_Supermarket_Sales'] = np.log1p(newdata['Product_Supermarket_Sales'])
sns.histplot(data=newdata['Product_Supermarket_Sales'], bins=15, ax=ax[1])

plt.tight_layout()
plt.title("Transformation of Product_Supermarket_Sales feature")

# next, let's transform Product_Shelf_Visibility
fig, ax = plt.subplots(1,2)

# plot of normal Product_Supermarket_Sales on the first axis
sns.histplot(data=newdata['Product_Shelf_Visibility'], bins=15, ax=ax[0])

# transform the Product_Supermarket_Sales and plot on the second axis
newdata['Product_Shelf_Visibility'] = np.log1p(newdata['Product_Shelf_Visibility'])
sns.histplot(data=newdata['Product_Shelf_Visibility'], bins=15, ax=ax[1])

plt.tight_layout()
plt.title("Transformation of Product_Shelf_Visibility feature")

# feature encoding
for col in cat_cols:
    print('Value Count for', col)
    print(newdata[col].value_counts())
    print("---------------------------")

# save the target value to a new variable
y_target = newdata['Product_Supermarket_Sales']
newdata.drop(['Product_Supermarket_Sales'], axis=1, inplace=True)

# one hot encode using pandas dummy() function
dummified_data = pd.get_dummies(newdata)
dummified_data.head()

# fill-in missing values
# print null columns
dummified_data.isnull().sum()

# compute the mean
mean_pw = dummified_data['Product_Weight'].mean()

# fill the missing values with calculated mean
dummified_data['Product_Weight'].fillna(mean_pw, inplace=True)

# check if filling is successful
dummified_data.isnull().sum()

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(dummified_data, y_target, test_size = 0.3)

print("Training data is", X_train.shape)
print("Training target is", y_train.shape)
print("test data is", X_test.shape)
print("test target is", y_test.shape)

from sklearn.preprocessing import RobustScaler, StandardScaler
scaler = RobustScaler()

scaler.fit(X_train)

X_train = scaler.transform(X_train) 
X_test = scaler.transform(X_test)

X_train[:5, :5]

from sklearn.metrics import mean_absolute_error
from sklearn.model_selection import KFold, cross_val_score


def cross_validate(model, nfolds, feats, targets):
    score = -1 * (cross_val_score(model, feats, targets, cv=nfolds, scoring='neg_mean_absolute_error'))
    return np.mean(score)

n_estimators=150
max_depth=3
max_features='sqrt'
min_samples_split=4
random_state=2

from sklearn.ensemble import GradientBoostingRegressor

gb_model = GradientBoostingRegressor(n_estimators=n_estimators, max_depth=max_depth, max_features=max_features, min_samples_split=min_samples_split, random_state=random_state)

mae_score = cross_validate(gb_model, 10, X_train, y_train)
print("MAE Score: ", mae_score)

from flytekitplugins.papermill import record_outputs
record_outputs(mae_score=float(mae_score))