This project has been rolled out with one of the world’s most successful volume carmakers that comprises multiple performance brands, and is considered to be a key player in the segment of smart mobility. The car-maker established a highly specialized research center that employs tens of researchers to keep growing in AI, augmented reality and climate-friendly alternatives. In doing so, it is paving the way into a responsible and sustainable start into the future of self-driving electromobility. Amongst others, company applications and business processes that require predictive analytics solutions are addressed in this research center.
Customers cancelling their services or upgrading to the product of a different company are a major challenge in any industry. A crucial task to make a business succeed is to define an adequate prediction model for customer churn: It is up to 25 times more expensive to acquire new customers than to retain old ones. Therefore, data scientists have to identify features that indicate potential customer churn to build accurate prediction models,. These models can identify churn on a personal customer level to enable individualized retention plans. As a result, marketing spendings get optimized while the client retains the maximum number of customers. The client’s research center has already developed such predictive models. Yet, the question remained, whether more accurate prediction models could be built in less time and whether future costs to build respective models could be reduced. Hence, this automotive manufacturer collaborated with the getML team to test automated feature engineering tools for relational customer relation management (CRM) data.
Business experts and data scientists alike have a strong interest in identifying as many influencing factors for customer churn as possible to build accurate prediction models. However, this is a manual and time-consuming procedure: There are many highly customized CRM applications that vary between markets, brands and transnational sales organizations. As a consequence, many different data sources result in an extremely complex data scheme that contains millions of potentially important features. Here, manual feature extraction is the main task to build ML algorithms that predict customer churn accurately. For every predictive analytics use case, it is therefore key to know where to start from. First, one has to define the term “churn”: Can one talk of “churn” already if a customer made use of the company's services just once? Second, do customers have to spend a certain amount of money, or do they have to make purchases at a specific rate to be worth the company’s effort to be retained? If so, how to define the threshold? Third, the period of inactivity must be determined to label customers properly as churned. Interaction rates, frequency rates of purchases or the period in between purchases are potential markers (also called features) for churn. Then, the company has to determine which of its churning customers will remain if targeted by tailor-made retention offers. The investment in customer loyalty programs is very expensive. Hence, these initiatives have to be well-developed and require predictions of very high accuracy. Furthermore, they must be easy to implement in transnational affiliated companies.
The getML team was confronted with multiple complex, market and brand-specific data schemes. These data schemes were collected out of multiple internal and external CRM systems. CRM systems, in turn, contain many different events such as customer touchpoints as well as market and brand-specific products and services. The customer churn prediction model of our client already worked quite accurately. However, in one market, getML outperformed the in-house-built model in less than 15 minutes of training time and increased the prediction accuracy by a further 4%. The getML Suite identified several market- and brand-specific features for customer churn that proved to be crucial for a higher predictive accuracy. In addition, these features delivered new insights into the motives behind customer churn. Based on these insights, the getML team made recommendations for an optimized customer service. The optimization helped to increase customer loyalty and to decrease retention rates. Moreover, the getML Suite significantly cut down the manual workload faced in predictive analytics. In fact, in one specific project, the getML Suite decreased the data science workload by ~90%: The getML team could prove that one data scientist who uses the getML Suite can achieve the results of five data scientists working on manual feature extraction and model building for two months. Hence, these use cases demonstrated not only a higher customer churn prediction accuracy and deeper insights into reasons behind retention. GetML also enabled our client to save economic resources substantially: Freed-up data science capacities can now be reinvested in the optimization of existing digital innovation products or in the exploration of new business cases. Overall, the performance quality of getML was so convincing that one of our client’s markets decided to use the getML Suite in its production environment.
The project collaboration with the getML team provided the client with a top-performing solution: In total, the getML Suite exceeded the in-house-built prediction system within 15 minutes of training time, and achieved a further 4% increase in customer churn prediction accuracy. Based on auto-generated features, the getML team gave insights into the underlying motives of customer retention. It made recommendations to increase customer loyalty and decrease customer churn. Moreover, getML cut down the workload of a data science project by ~90%. The client can invest the resources saved by automated feature engineering in new business cases and in the development of further digital innovation products. Therefore, the getML Suite continues to be the tool of choice for this automotive manufacturer, to build and deploy customer churn prediction models. Finally, the getML Suite continues to provide accurate customer churn scores in a live environment in one of the client’s core markets.
Predict item purchases that will be gifts. This analysis is based on a public domain data set provided by the American Bureau of Labor Statistics.
Automate Machine Learning on time series with getML