Improving prediction of customer behavior in nonstationary environments

Customer churn, switching from one service provider to another, costs the wireless telecommunications industry $4 billion each year in North America and Europe. To proactively build lasting relationships with customers, it is thus crucial to predict customer behavior. Machine learning has been applied to churn prediction, using historical data such as usage, billing, customer service, and demographics. However, because customer behavior is often nonstationary, training a model based on data extracted from a window of time in the past yields poor performance on the present. We propose two distinct approaches, using more historical data or new, unlabeled data, to improve the results for this real-world, large-scale, nonstationary problem. A new ensemble classification method, with combination weights learned from both labeled and unlabeled data, is also proposed, and it outperforms bagging and mixture of experts.

1 Introduction

Customer churn, switching from one service provider to another, destroys profits and decreases shareholder value. In the wireless telecommunications industry, the annual churn rate reaches 25% in Europe and 30% in the United States [1]. It costs around five times as much to sign on a new subscriber as to retain an existing one. In most developed markets, acquiring a new customer costs an average of $300 to $400. It is estimated that, in the mature markets of North America and Europe, churn costs wireless service providers a combined total of more than $4 billion each year [1]. It is thus crucial to predict customer behavior, e.g. churn, in advance. Accurate prediction may allow one to forestall churn by proactively building lasting relationships with customers.