100 Celsius AI

Predictive customer analytics

How Machine Learning can help you manage user retention

Let's imagine you are an employee in a medium-sized telecommunications company. You are running over 100 marketing campaigns per month, most of them focused on customer retention. How do you identify and address customers with a high risk of terminating their contract?


Typically you might work off your experience. As a first step, you may try to identify profiles of your customers which have terminated in the past using a mixture of your intuition and the data you have at hand. This way you would probably identify that customers on contract X have a high risk especially in the high revenue segment. Great! Now do this for the up to thousands of price plans and for hundreds of variables describing your customer and his usage behaviour. Feeling overwhelmed? 


Enter Machine Learning for data-driven churn management.


Machine Learning, especially supervised Machine Learning helps you identify hidden correlations in the data that would otherwise be impossible to spot. It is supervised because we feed in historic information on customer that have churned and retained. Many Machine Learning algorithms are available which work well on supervised data. But since customer usage data is time-dependent (i.e. how many GB's a customer has used in a particular month/day or even in a given session), many standard Machine Learning algorithm such as logistic regression or decision trees do not work on the raw data.


There are multiple approaches to remedy this:

  1. Data conversion
  2. Tailored Machine Learning algorithm

The more straight-forward one is to convert the data into a format which standard algorithms are able to process. In this approach, the time series is aggregated over a fixed or variable amount of time. For example the data usage on the day before the data cutoff (i.e. the date of the data-dump) or the day/week/month the user terminated his contract. The duration over which is aggregated and how many buckets are considered is then a parameter of the model which is usually tuned via cross-validation. 

Fig1. - For many machine-learning algorithms, time series need to be aggregated, the length of the "aggregation-window" is a parameter to be tuned.


This can be a time-consuming process and must potentially be repeated every-time a new model is fitted.

A more general approach is to choose an algorithm which is designed for sequential data such as time series, for example Recurrent Neural Networks (RNN's) or Hidden Markov Models. While the aggregation method is easier to interpret and understand, the latter is generally more accurate. Engineering these models and retraining them due to changes in user behaviour and or contracts plus making it accessible to design campaigns is no small undertaking. 


At 100 Celsius AI, we use an ensemble of both wide (i.e. logistic regression) and deep models (RNN's) to power our easy-to-use campaign designer, giving powerful Machine Learning to the hands of marketing staff and combining the best of both methods. Talk to us about how we can help you to leverage the power of AI for maximizing retention in your company.


In a future blogpost we'll describe how we optimize campaign targeting based on continuous feedback to help our customers quickly improving their retention.





April 7, 2017


Pascal de Buren

Share This Post