EXPRESS: Modeling Dynamic Consumer Preferences from Few-shot Data: A Meta-Learning Approach

The ability to quickly capture and adapt to customer preferences is central for firms seeking to offer personalized products and improve retention. This objective becomes challenging when individual-level data on customer interactions are limited, as is often the case for new customers or short consumption sessions. To this end, we propose meta-temporal processes (MetaTP), a meta-learning framework that enables scalable personalization from a small number of individual observations. MetaTP is trained across a large collection of session-based tasks, allowing it to improve data efficiency and transfer shared structure across customers. To model customer interactions over time, MetaTP integrates a Transformer-based architecture that captures sequential consumption patterns within sessions. This design uncovers dynamic preference heterogeneity and enables accurate predictions. We illustrate MetaTP through an application on customer sequential consumption of digital products, focusing on the lukewarm stage of the customer journey, a transition period characterized by limited individual observations. Empirically, MetaTP outperforms a comprehensive set of benchmark methods in few-shot prediction and reveals meaningful patterns of preference evolution through its interpretable parameters. Managerially, we demonstrate how firms can leverage MetaTP to optimize personalized recommendations with limited individual data, including product sequencing decisions and both open-loop and closed-loop session completion strategies.