This post presents a snippet of a framework for defining the business problem in a systematic way using example.
The Problem in General Terms #
A large telecommunications operator is having a major problem with customer retention in their mobile telecom business. 8% of mobile telecom customers leave when their contracts expire, and it is getting increasingly difficult to acquire new customers as the mobile telecom market is now saturated.
The Approach [1] #
1. Identify the business problem
8% of mobile telecom customers leave when their contracts expire
2. Determine why machine learning is the most appropriate approach
See When to Apply Machine Learning
3. List possible solutions and define key items
Solution 1: Predict churn
Proposed model: Build a model that predicts the likelihood of a customer to leave
Usage: Run the model every month to identify customers most likely to leave, then offer them an incentive to stay
Benefit: Retaining existing customers is much less expensive than attracting new ones. Average Revenue Per User after offering prospective churners would be $2.09 if 20% of prospective churners are retained but would fall to $1.83 if all prospective churners actually churn.
Data requirements: a large collection of historical data marked as churn and non-churn for each customer; demographic, behavioral and transactional information about each customer, information about the company’s services
Capacity requirements: capacity to build suitable incentives, a mechanism to contact identified customers with the incentive (e.g., by email, text messaging, voice call, etc.)
Business success: Reduce churn by 20% in the first 3 months of deployment
Business failure: Unable to reduce churn by 18% in the first 6 months of deployment
Machine learning success: Precision of 0.85
Machine learning failure: Precision of 0.7 or less
Solution 2: Identify reasons for leaving
Proposed model: Identify a small set of features of the company’s products that are important in building a model that predicts the likelihood of a customer to leave
Define the rest of the items in the manner done in Solution 1 above
Solution 3: Design incentives
Proposed model: Build a next-best-offer model that accurately predicts the likely effectiveness of incentives that could be offered to entice customers with high churn-likelihood to stay
Define the rest of the items in the manner done in Solution 1 above
Solution 4: Recommend a service
Proposed model: Build a model that accurately predicts how much a customer would like an existing service
Define the rest of the items in the manner done in Solution 1 above
4. Prioritize models
Draw up an implementation priority list according to the feasibility of the above models in terms of data and capacity requirements
5. Prepare an initial implementation plan
Develop a high-level implementation plan that can be iterated on as more information is discovered including project schedule, tools and techniques, risk assessment, and resource allocation
References #
- Inspired by Kelleher, J. D., Mac Namee, B., and DArcy, A. (2015). Fundamentals of machine learning for predictive data analytics : algorithms, worked examples, and case studies. Massachusetts Institute of Technology, 64-70
