12:00 Meta Ensembles in Marketing: Learning across Models and Tasks
Aurelie Lemmens, Erasmus University Rotterdam
Making predictions is an essential task of marketing research. While most studies investigate the performance of prediction models on a rather limited number of empirical data sets (usually one or two), the No Free Lunch theorem proposed in the machine learning literature warns us about the variability of a model’s performance across data contexts. The goal of this article is to raise awareness about the lack of generalization of a method performance across applications but also to provide some good vibes in the form of a viable solution to this problem. To do so, we assembled a collection of 25 marketing data sets that cover different binary predictions tasks and different industries and implemented a broad set of binary prediction algorithms based on past literature. Our analysis confirms a high degree of variability in performance across data contexts. To address this problem, we then propose a new algorithm that combines the benefits of ensemble learning (i.e., combining models within a given task) and meta learning (i.e., learning across tasks), our “good vibe” contribution. Our approach significantly reduces the variability in performance across tasks and leads to a higher overall predictive performance than existing ensemble methods. We hope our work encourages researchers to consider the generalizability of their work across data contexts and fosters the adoption of meta learning in marketing.