This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Contents Business Understanding: Introduction ......................................................................................................... 2
Business Objective ........................................................................................................................................ 2
Data Mining Objective .................................................................................................................................. 2
Data Set ......................................................................................................................................................... 2
Data Preparation ........................................................................................................................................... 2
Data Modeling ............................................................................................................................................... 3
1. Decision Tree (Binary) ....................................................................................................................... 3
2. Decision Tree (Three-way tree) ........................................................................................................ 5
Profitability of a Proactive Retention Plan .................................................................................................... 9
The key variables predicting churn: ............................................................................................................ 10
Possible Incentives Offered ........................................................................................................................ 10
Test Measures ............................................................................................................................................. 11
Net Additions minus Existing Customers ................................................................................................ 11
1 | P a g e
List of Figures
Figure 1: Process flow Diagram ..................................................................................................................... 3
Figure 2: 2-way decision tree that resulted from cell2cell data set ............................................................. 4
Figure 3: Variables in descending order of their importance helping in splits for 2-way Tree .................... 4
Figure 4: Result summary of 2-way decision tree ......................................................................................... 4
Figure 5: Variables in descending order of their importance helping in splits for 3-way Tree .................... 5
Figure 6: Result summary of 3-way decision tree ......................................................................................... 5
Figure 7: Result summary of Regression without transformation variables ................................................ 5
Figure 8: Result summary of Regression with transformation variables ...................................................... 6
Figure 9: Result summary of Neural Network without Variable transformation and selection ................... 6
Figure 10: Result summary of Network with Variable transformation and selection .................................. 7
Figure 11: Comparison of cumulative lift value for different techniques ..................................................... 8
2 | P a g e
Cell2Cell: The Churn Game
Business Understanding: Introduction Cell2Cell is the 6th largest wireless company in the US, giving service to nearly 10 million subscribers,
serving more than 210 metropolitan markets & 2900 cities (covering nearly all 50 states). The company is
currently facing a major problem of customer churn.
We are using SAS EM 4.3 to develop a model for predicting customer churn at Cell2Cell.
Business Objective Reduce churn for the company
Improve profitability
Identifying incentives offered to the customers with high risk of churning
Data Mining Objective To develop an accurate predictive churn model (Lift value of at least 1.75)
To identify the factors that are important in driving subscribers churning
Data Set The given data set consists of 71,047 rows & containing a total of 78 variables (including a variable named
“CHURN”, signifying whether the customer had left the company two months after observation). One of
the variables named “CALIBRAT” was used to differentiate the validation dataset from training dataset.
Training dataset contained data of 40,000 customers and validation dataset contained 31,047 customers.
Data Preparation The dataset was divided in training and validation datasets, using “CALIBRAT” as the partition variable
(value of 1 was used training and value of 0 was used for validation).
Variable “CHURN” was set as target variable and some other variables (those not related to business
objective) were rejected.
Variable No.
Original Variable Changed Variable
22 Churn Target
26 CSA Rejected
30 Customer Rejected
77 Calibrat Rejected
3 | P a g e
78 Churndep Rejected
Table1. Showing variables that were rejected
Data Modeling Total of 6 different models were used to predict the churn of customers. These models were:
Decision Tree (binary)
Decision Tree (three way tree)
Logistic Regression
Logistic Regression with Transform Variables
Neural Networks
Neural Networks after transform variables and variable selection
SAS EM 4.3 was used to run these 6 models. Snapshot of the model is shown below.
Figure 1: Process flow Diagram
1. Decision Tree (Binary) For both 2-way and 3-way tree gini-reduction method was used. The assessment criteria was set to be
“Proportion of event in top 10%”.
4 | P a g e
Figure 2: 2-way decision tree that resulted from cell2cell data set
Figure 3: Variables in descending order of their importance helping in splits for 2-way Tree
As can be seen from above figure, EQPDAYS – Number of days of the current equipment (split at <302),
MONTHS – Months in service (Split < 11 months) are most important variables that resulted in splits.
Figure 4: Result summary of 2-way decision tree
As can be seen from figure above, with number of leaves greater than 34, no significant split happens.
5 | P a g e
2. Decision Tree (Three-way tree)
Figure 5: Variables in descending order of their importance helping in splits for 3-way Tree
The important variables are very similar to that used in 2-way decision tree. First four variables are same
for 2-way and 3-way tree.
Figure 6: Result summary of 3-way decision tree
With number of leaves greater than 93, no significant split happens in 3-way tree.
3. Logistic Regression Here, no transformation of variable was carried out.
Figure 7: Result summary of Regression without transformation variables
6 | P a g e
4. Logistic Regression with Transform Variables In this model, few variables were transformed. Details of transformation of variables are as below:
1. Variable MOU- Minutes of Usage was transformed using log transformation. The variable had high
skewness earlier.
2. CUSTCARE- Mean number of customer care calls was transformed into a squared variable. As high
number of customer care calls suggests high number of complaints, it can be a major cause for
churn. Sqauring the variable helps increase its influence.
3. Using decision trees, the 2 major variables EQPDAYS and MONTHS were identified and
transformed by creating 2 buckets. For EQPDAYS, the cut off value for bucket used is 301 days
and for MONTHS, the cut off value used is 10 months.
4. Other important variables like CHANGEM and CHANGER were transformed but they didn’t
improve performance.
Figure 8: Result summary of Regression with transformation variables
5. Neural Networks Here, no transformation and variable selection was done.
Figure 9: Result summary of Neural Network without Variable transformation and selection
7 | P a g e
6. Neural Networks after transform variables and variable selection Here, the transformed variables were made to pass through variable selection process with default
settings. Initially, transformed variable CUSTCARE was rejected by neural network model. But, the same
was forced to be used. With this the performance of the neural network model improved.
Figure 10: Result summary of Network with Variable transformation and selection
As can be seen from figure above, the average error has come down a little bit and has less variation as
compared to that of neural network model without transformation and variable selection.
8 | P a g e
Evaluation The lift chart of all the techniques is shown below:
Figure 11: Comparison of cumulative lift value for different techniques
Legend:
Tree : 2-way decision tree
Tree-2 : 3-way decision tree
Neural : Neural Network without transformations
Neural-2 : Neural Network with transformed variables and variable selection
Reg : Regression without variable transformation
Reg-2: Regression with variable transformations
Table below shows cumulative lift values for different techniques at different percentiles (deciles).
Table 2. Showing cumulative lift values at different deciles for different techniques
As can be seen from table, the performance of “Regression with transformed variables” is best among the
different techniques used. The lift value at first deciles with Regression with transformed variables is 2.102
Percentile 2-way Tree 3-way Tree Regression
Regression with
Transformation
Neural
Networks
Neural Networks
with
Transformation
10 1.931 1.904 1.658 2.102 1.511 1.691
20 1.706 1.699 1.642 1.79 1.42 1.593
30 1.451 1.573 1.5 1.604 1.336 1.489
40 1.353 1.444 1.4 1.457 1.318 1.416
50 1.294 1.325 1.343 1.411 1.297 1.346
60 1.254 1.263 1.253 1.314 1.21 1.284
70 1.236 1.192 1.203 1.217 1.159 1.194
80 1.127 1.131 1.131 1.133 1.127 1.131
90 1.056 1.07 1.067 1.071 1.076 1.062
100 1 1 1 1 1 1
9 | P a g e
which is highest. It is also evident from table that “Regression with transformed variables” performs best
at most deciles and significantly better till top 3 deciles.
The performance of Neural Network technique is worse among the 3 techniques. Transformation and
variable selection does help in improving the performance of Neural Network technique.
3-way decision tree, though underperforms when compared to 2-way decision tree at first deciles,
performs better than 2-way decision tree at other deciles.
So, based on the table above, the best technique is “Regression with transformed variables” which gives
a lift value of 2.102 which is well above the target of 1.75.
Profitability of a Proactive Retention Plan Using regression model with transformed variables as inputs, the following values are calculated.
Assumption: Subscriber in the 1st deciles is targeted.
β = Base line churn rate= 1.96 %
λ= Lift = 2.102
γ = Success rate = λ-1= 1.102
LVC= Lifetime value of customer
C= Cost of incentive
Profit = Probability of Churner* Success Rate* LVC-C