codejack-cr / social-network-ads

A small project using Keras and Keras Auto-Tuner API to build a logistic regression model to predict whether a user will purchase an item or not based on historic user data on Age, Gender, Salary and Purchase Status

Social Network Ad based Predictions

Setting up the environment

import pandas as pd
df = pd.read_csv('../Social-Network-Ads/Social_Network_Ads.csv')
# check for number of null entries
print('Data is packed' if not df.isnull().values.any() else 'Data contains missing values')
df.head()

Data is packed

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	User ID	Gender	Age	EstimatedSalary	Purchased
0	15624510	Male	19	19000	0
1	15810944	Male	35	20000	0
2	15668575	Female	26	43000	0
3	15603246	Female	27	57000	0
4	15804002	Male	19	76000	0

Visualizing the data

Heatmap

import seaborn as sb
import matplotlib.pyplot as plt

sb.heatmap(df.corr(), annot = True, cmap = 'coolwarm', vmin = -1, vmax = 1)
plt.show()

Interpretable graph

from matplotlib.colors import ListedColormap

# Age vs salary. Men are blue points, Females are Red points and No Buyers are black points
salaries = df['EstimatedSalary']
ages = df['Age']
genders = df['Gender']
buy_stats = df['Purchased']
colors = [1 if genders[i] == 'Male' and buy_stats[i] == 1 else 2 if genders[i] == 'Female' and buy_stats[i] == 1 else 0 for i in range(len(df['User ID'].tolist()))]
cmap = ListedColormap(['black','blue','red'])
classes = ['Not Bought', 'Male', 'Female']
plt.clf()

fig = plt.figure()
fig.suptitle("Impact of salary and age")
scatter = plt.scatter(salaries, ages, c = colors, cmap = cmap)
plt.legend(handles=scatter.legend_elements()[0], labels=classes)
plt.xlabel('Salary-->')
plt.ylabel('Age-->')
plt.show()

<Figure size 432x288 with 0 Axes>

Preprocessing the data

Data Cleaning and Scaling

# User ID is irrelevant
df = df.drop(['User ID'], axis = 1)

# We can change Male and Female to 0 or 1
df['Gender'] = df['Gender'].apply(lambda gender: 0 if gender == 'Male' else 1)

# Normalize Age and EstimatedSalary based on mean and range
# value = (val - mean)/(max - min)

# Salary Normalization
import math
salaries = df['EstimatedSalary']
mean_salary = min(salaries) + max(salaries) / len(salaries)
salary_range = max(salaries) - min(salaries)
norm_salaries = [(salary - mean_salary) / salary_range for salary in salaries]

# Age normalization
ages = df['Age']
mean_age = min(ages) + max(ages) / len(ages)
age_range = max(ages) - min(ages)
norm_ages = [(age - mean_age) / age_range for age in ages]

# Assign to df
df['Age'] = norm_ages
df['EstimatedSalary'] = norm_salaries
df.head()

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	Gender	Age	EstimatedSalary	Purchased
0	0	0.020238	0.026852	0
1	0	0.401190	0.034259	0
2	1	0.186905	0.204630	0
3	1	0.210714	0.308333	0
4	0	0.020238	0.449074	0

Train and Test split

# Use 60% data as training data and 20% as validation data and 20% as test data
y = df.pop('Purchased').to_numpy()
X = df.to_numpy()

# Split
from sklearn.model_selection import train_test_split

# Train+Validation and Test split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2)
# Train and Validation split
X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size = 0.25)

print(f'''Training size: {len(X_train)}
Validation size: {len(X_valid)}
Test size: {len(X_test)}''')

Training size: 240
Validation size: 80
Test size: 80

Creating Keras Model

from tensorflow import keras

#ANN
model = keras.Sequential()
model.add(keras.layers.Flatten(input_shape=(X_train.shape[1], )))
model.add(keras.layers.Dense(int(X_train.shape[1]*2), input_dim=X_train.shape[1], activation='relu'))
model.add(keras.layers.Dense(X_train.shape[1]*2, activation = 'relu'))
model.add(keras.layers.Dense(X_train.shape[1]*2, activation = 'relu'))
model.add(keras.layers.Dense(1, activation = 'sigmoid'))
model.compile(optimizer = 'adam',
             loss = 'binary_crossentropy',
             metrics = ['accuracy'])
print(model.summary())

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten (Flatten)            (None, 3)                 0         
_________________________________________________________________
dense (Dense)                (None, 6)                 24        
_________________________________________________________________
dense_1 (Dense)              (None, 6)                 42        
_________________________________________________________________
dense_2 (Dense)              (None, 6)                 42        
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 7         
=================================================================
Total params: 115
Trainable params: 115
Non-trainable params: 0
_________________________________________________________________
None

#Set epochs
epochs = 100

#Clear output
import IPython
class ClearTrainingOutput(keras.callbacks.Callback):
  def on_train_end(*args, **kwargs):
    IPython.display.clear_output()
    
#Fit the model
trained_model = model.fit(X_train, y_train, epochs = epochs, validation_data = (X_valid, y_valid), callbacks = [ClearTrainingOutput()])

print('Max Accuracy Reached: {0:0.2f} %'.format((max(trained_model.history['accuracy']) * 100)) )
# Print the learning curve
plt.plot(range(1, epochs+1), trained_model.history['loss'] , 'r', range(1, epochs+1), trained_model.history['val_loss'], 'b')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training Loss', 'Validation Loss'])

Max Accuracy Reached: 84.17 %





<matplotlib.legend.Legend at 0x7f320b9c9ca0>

Making Predictions

predictions = model.predict(X_test)
from sklearn.metrics import accuracy_score
predictions = [0 if prob < 0.5 else 1 for prob in predictions]
print("Test Set Accuracy: {0:0.2f} %".format(accuracy_score(y_test, predictions) * 100))

Test Set Accuracy: 85.00 %

Creating auto-tuned Keras model

Defining model with variable hyperparameters

from tensorflow import keras
import kerastuner as kt
def model_builder(hp):
    model = keras.Sequential()
    model.add(keras.layers.Flatten(input_shape=(X_train.shape[1], )))
    for i in range(hp.Int('layers', 1, 10)):
        model.add(keras.layers.Dense(
            units=hp.Int('units_' + str(i), X_train.shape[1], X_train.shape[1]*5, step = X_train.shape[1]),
            activation='relu'))
    model.add(keras.layers.Dense(1, activation='sigmoid'))
        
    model.compile(optimizer = 'adam',
                loss = 'binary_crossentropy', 
                metrics = ['accuracy'])
    return model

tuner = kt.Hyperband(model_builder,
                     objective = 'val_accuracy', 
                     max_epochs = 20,
                     factor = X_train.shape[1],
                     directory = 'auto_tuned_model',
                     project_name = 'soc_net_ad_pred')

tuner.search(X_train, y_train, epochs = 20, validation_data = (X_valid, y_valid), callbacks = [ClearTrainingOutput()])
best_hps = tuner.get_best_hyperparameters(num_trials = 1)[0]

Trial 30 Complete [00h 00m 02s]
val_accuracy: 0.8125

Best val_accuracy So Far: 0.8125
Total elapsed time: 00h 00m 51s
INFO:tensorflow:Oracle triggered exit

## Create model from tuned hypermodel
model = tuner.hypermodel.build(best_hps)
print(model.summary())

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten (Flatten)            (None, 3)                 0         
_________________________________________________________________
dense (Dense)                (None, 12)                48        
_________________________________________________________________
dense_1 (Dense)              (None, 15)                195       
_________________________________________________________________
dense_2 (Dense)              (None, 15)                240       
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 16        
=================================================================
Total params: 499
Trainable params: 499
Non-trainable params: 0
_________________________________________________________________
None

Train the model with best hyper parameters

epochs = 100
trained_model = model.fit(X_train, y_train, epochs = epochs, validation_data = (X_valid, y_valid), callbacks = [ClearTrainingOutput()])

print('Max Accuracy Reached: {0:0.2f} %'.format((max(trained_model.history['accuracy']) * 100)) )
# Print the learning curve
plt.plot(range(1, epochs+1), trained_model.history['loss'] , 'r', range(1, epochs+1), trained_model.history['val_loss'], 'b')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training Loss', 'Validation Loss'])

Max Accuracy Reached: 87.92 %





<matplotlib.legend.Legend at 0x7f31c4137790>

Making predictions with auto-tuned model

predictions = model.predict(X_test)
from sklearn.metrics import accuracy_score
predictions = [0 if prob < 0.5 else 1 for prob in predictions]
print("Test Set Accuracy: {0:0.2f} %".format(accuracy_score(y_test, predictions) * 100))

Test Set Accuracy: 91.25 %