spy16 / churn-prediction

Customer churn prediction using Deep Neural Network.

Churn Prediction

Churn prediction exercise using Deep Neural Network.

Problem

Given various information about current and exited customers of a certain bank, predict wether a current customer is likely to churn.

Customer data is given as a CSV file: customer-data.csv

Solution

Solution is based on a deep neural network that outputs the probability of churn (a scalar).

1. predict_churn_v1.py:

   • Architecture: (11) ==> (6, ReLU) ==> (6, ReLU) ==> (1, Sigmoid)
   • Hyper parameter tuning: manual
   • Accuracy (train): 84.09%
   • Accuracy (test) : 82.13%

2. predict_churn_v2.py:

   • Architecture: (11) ==> Dense(6, ReLU) ==> Dropout@0.2 ==> (6, ReLU) ==> Dropout@0.2 ==> (1, Sigmoid)
   • Hyper parameter tuning: manual
   • Accuracy (train): 82.92857142857143%
   • Accuracy (test) : 82.23333333333333%

3. predict_churn_v3.py:

   • Architecture: (11) ==> Dense(6, ReLU) ==> Dropout@0.2 ==> (6, ReLU) ==> Dropout@0.2 ==> (1, Sigmoid)
   • Hyper parameter tuning: manual
   • k-Fold Cross Validation: Max=0.860000 Min=0.774286 [Mean=0.802714 Variance=0.022619]
   • Accuracy (train): 79.27142857142857%
   • Accuracy (test) : 80.46666666666667%

4. predict_churn_v4.py:

   • Architecture: (11) ==> Dense(6, ReLU) ==> Dropout@0.2 ==> (6, ReLU) ==> Dropout@0.2 ==> (1, Sigmoid)
   • Hyper parameter tuning: using GridSearchCV
   • Best Accuracy: 81.9571%
   • Best Params
     • batch_size=25 (Tested for 25 & 32)
     • epochs=500 (Tested for 100 & 500)
     • optimizer=rmsprop (Tested for rmsprop and adam)