Cell2Cell Churn Modeling

In the project, my team developed a model to predict customer churn for Cell2Cell - a telecommunications service provider. The collected data is outdated, considering that the telecommunication market was reaching a saturation point and competition was rapidly increasing. It is of utmost importance to develop a model to identify the most important drivers of churn.

Let me give you a brief introduction of our methodology:

Dataset -

• 71,047 observations and 78 variables.

• “Calibration” data: 40,000 customers (Calibrate = 1)

• “Validation” data: 31,047 customers (Calibrate = 0)

Data Cleaning and Feature Selection (With Python/SQL) :

• Imputed missing values of continuous variables with the calculated mean of a specific column. (Worked better than KNN Imputation)

• Removed variables with high correlation (>70%)

• Removed “csa” due to high number of levels (n=744) and customer ID (Worked better than recording area codes to “US state” dummies)

• Employed SelectKBest feature selection method which calculates a said score for each of the variables depending on how they influence the target variable. By default considers the ANOVA F scores and ranks the variables based on the scores.

We selected the Top 10 important variables for the following modeling.

Data Modeling (With R/Machine Learning)

• Applied 4 machine learning algorithms - Random, Logistic Regression, Random Forest, Gradient Boosting to find out the most suitable one

• Estimated models by Height of the lift curve and AUC

After finding the best model, we divided records into three-level churn groups and calculated their CLV to the company, which gave us a much clearer answer that customers with lower churn rate would bring higher values.

Here is an explanation of valuables that have a positive effect on customer churn.

With these new insights, we recommended a pro-active customer churn management program.