Foundation Models in Marketing
(A2025-125960)

AUTHORS

Daniel Ringel, University of North Carolina at Chapel Hill; Florian Ellsaesser, University of Bristol; Sebastian Gabel, Rotterdam School of Management; P. K. Kannan, University of Maryland; Xiao Liu, New York University (NYU) - Leonard N. Stern School of Business

ABSTRACT

Foundation models are large, pretrained machine learning models designed to generalize across various tasks by learning patterns from extensive input data. The objective of this special session is to explore how foundation models, when pretrained or adapted with marketing-specific data, can address complex challenges in marketing. Across four research papers, we demonstrate the benefits of advanced machine learning architectures in improving predictive accuracy, interpretability, and strategic decision-making across various marketing domains. “Interpretable Recommendations and User-Centric Explanations with Geometric Deep Learning” by Yan Leng, Xiao Liu, and Rodrigo Ruiz addresses challenges in recommender systems such as sparse ratings and cold-start problems. The authors propose a two-stage framework using a Multi-Graph Graph Attention Network (MG-GAT) to integrate multimodal data, including network information, attributes, and ratings. Using Yelp data, they demonstrate that their approach improves predictive accuracy over existing algorithms and enhances user acceptance, satisfaction, and trust in a randomized user experiment. “Leveraging External Data to Inform Customer Acquisition Decisions” by Florian Ellsaesser, Sebastian Gabel, Tetyana Kosyakova, and Aurelie Lemmens investigates whether customers’ transaction data from other companies can predict their future behavior at a focal company. The authors propose a methodological framework that leverages Long Short-Term Memory neural networks (LSTMs) to process external bank transaction data. They validate their approach using bank transaction data to predict cross-retailer transactions and various post-acquisition behaviors. “GenAI for Customer Journeys: A Transformer Approach” by Zipei Lu and P.K. Kannan applies transformer-based models to analyze large numbers of customer interactions across multiple channels. By treating customer interactions as sequences analogous to words in a sentence, the model uses self-attention mechanisms to capture dependencies among interactions. This approach allows for the prediction of visit and purchase probabilities, as well as sequence-level classification tasks like predicting customer churn or lifetime value. “The Market Basket Transformer: A New Foundation Model for Retail” by Sebastian Gabel and Daniel M. Ringel adapts transformer architectures to retail analytics to handle high-dimensional and sparse retail data. The authors implement a customized transformer model tailored to market basket data, capturing complex product relationships and shopping patterns. They address data sparsity issues, particularly for long-tail products, with a more efficient training approach and by generating additional training examples. They further demonstrate that their pretrained market basket transformer can be fine-tuned for specific tasks like predicting coupon redemptions with minimal additional training data. Collectively, these four papers highlight the growing relevance of foundation models in marketing. Each paper connects foundation models—be it pretrained or adapted—with marketing-specific data to enhance predictive accuracy, interpretability, and strategic decision-making. The session offers insights into the practical implementation of foundation models and highlights their potential to transform current marketing practices.