Data Science Nugget - 2

This is the second article in this series and will cover the application of Data Science to Churn Management.

Churn management covers the identification of customers who are likely to churn from a company and identifying the proactive steps that need to be taken to retain them. Another important to be considered in this context is the value the customer brings to the company. In case of Telcos, there are a large number of customers who generate lesser revenue than the cost to service them. In such cases, it will be profitable for the Telco to let such customers churn out.

Churn prediction process identifies the attributes of the customer that can can be an indicator of a potential churner. A homogeneous data set of past churners based on the business context is created using clustering algorithms or simple filters on the data set. The business context here could be age group, geographical location, customer segment etc. Churn prediction is a classification problem where the desired output is a discrete label (Yes or No)

A churn data model is created using the identified attributes with the flag for churn identified as 'YES" and the data model is trained using different classification algorithms like Decision tree, Logistic regression, Random Forest etc.

• The identified Data set is randomly partitioned into four data sets - one for training and three for testing purposes. The data sets for training and testing can be used interchangeably.
• Ideally we can train the Data model with three algorithms with the training set data.
• The learning algorithm finds patterns in the training data that map the input data attributes to the output attribute
• The learning algorithm will help identify the principal attributes that have the maximum impact of the output attribute.
• This Attribute reduction is an important step in improving the efficiency and the quality of data mining.
• The trained Data model is tested using the three sets of test data.
• Lift Charts and Confusion matrix are created for each execution and the accuracy of the model/algorithm is documented. The best model/algorithm is selected and deployed in production.
• The model/algorithm is tested and trained using fresh data sets at periodic intervals to ensure the reliability and accuracy of the predictions.

Any Churn reduction actions can be made more relevant by including the revenue and the profitability of the customer as a key consideration in the preparation of the Data set.

It should also be noted that while Black box algorithms like Neural networks provide greater accuracy, White box algorithms like Decision tree are easier to understand and interpret.