Introduction to Statistical and Machine Learning Methods for Data Science by Carlos Andre Reis Pinheiro & Mike Patetta

Author:Carlos Andre Reis Pinheiro & Mike Patetta [Pinheiro, Carlos Andre Reis & Patetta, Mike]
Language: eng
Format: epub, pdf
Publisher: SAS Institute

Use Case Study: Churn Model for Telecommunications

An example of where ensemble models might be especially useful is in churn in telecommunications. Telecommunications companies usually have millions of customers, and customers might present several distinct types of behavior of usage, payment, and even making churn. Eventually, a single model would not be able to capture all of these different relationships between the input variable and the target for subgroups of customers. Of course, one option to mitigate this is to first create a customer segmentation (clustering) and then create one churn model for each cluster. However, even inside the same cluster, we can observe different customer behaviors. An ensemble model can work well in capturing all these different behaviors and correlating them to the target, in this case the churn event.

Even though we have the ensemble model as the final model, we also have the individual models that contributed to the ensemble approach. As mentioned before, the ensemble model lacks interpretability. However, data scientists can use the ensemble model as the decision in a retention campaign (offer something to the customer or not) and can also use some of the individual models (the one with the highest predictive probability) to eventually explain the result or a characteristic associated to the churn event. Data scientists need to understand that business departments require more than just a probability. They also want to understand the reasons customers are making churn, the reasons customers are not going to pay their bills, and the reasons customers are going to purchase a product. All these reasons help business departments create the marketing campaigns, the offerings, the marketing messages, and the channels to contact the customers. A minimum of interpretability is needed to support a full business action. Therefore, the combination of the ensemble model to select the customers and the individual models to interpret the results can extract most of the benefits from analytical models.

A churn model quite often uses all the information about the customer, considering the time frame perspective. How the customers are changing their behavior right before the churn, and how these characteristics are correlated to the target, or in this case the churn event. As shown in Figure 5.7, common input variables in churn models include:

Demographic variables: customer characteristics such as age, gender, and education

Usage variables: how much usage of products and services and indicators over time to characterize the trends of usage

Product/service portfolio: what bundles and packages the customer has and for how long

Revenue behavior: how much revenue the customer generates, for how long, and indicators over time to characterize trends about the average revenue per user

Payment behavior: how many defaults the customer has had in the past and the length of the delay

Contract information: how much time before the contract ends

Contacts in the call center: the frequency and recency of contacts to the call center, including the reasons and the resolutions

Defects reported: the frequency and recency of defects reported by the customer, the types, the time to fix them, and the

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.