Business-Oriented Data Analytics: Advances in Profit-Driven Model Building and Fraud Detection

In recent years, the ever-larger becoming avalanche of data poses a new set of challenges for businesses. Yet, where there is risk, there also lies opportunity!  Consequently, businesses rely more than ever on “smart” algorithms to stay ahead of their competition.

As for the mature market of a mobile phone service provider, for example, it is vital to the provider’s business to prevent its customers from leaving the company (i.e., from churning); especially, since research studies showed that acquiring a new customer is much more expensive than to retain an existing one.  Leveraging the information stored in its massive databases, the telephone company (telco) wishes to identify in advance the customers who are about to leave in order to retain them through an appropriate action. The telco, however, is not interested in preventing every customer from churning, rather its primary focus, due to limited resources, is to first retain those customers that are the most valuable to the business.

Hence, the dissertation presents a set of profit-driven classifiers that account for the benefits and costs of a business action and optimize for maximum profit when learning from data. In this way, we show that the profit-driven classifiers first identify churners that have a higher value to the business.

As telcos, insurance companies are also in need of “smart” algorithms to leverage their data as, for example, for the detection of fraud. Hence, this dissertation additionally presents an anomaly detection algorithm to support insurance companies to close their value leaks caused by fraud. More specifically, in collaboration with a large European insurer, we present a real-world case study on workers’ compensation fraud, showcasing the application of the new anomaly detection algorithm. We demonstrate the construction of a meaningful anomaly score that allows for the detection of conditional anomalies.

In summary, this doctoral dissertation offers a set of new machine learning algorithms developed in the context of important business applications that have a real-world implication. At the core of their development was the sharp focus on the creation of value for businesses.