PREFER: Personalized Review Summarization with Online Preference Learning

Product reviews are crucial for online shopping, but the sheer volume can overwhelm buyers. Current e-commerce summarization tools produce generic, static summaries that ignore the fact that different users prioritize different product features—and those priorities change over time. To solve this, researchers Millend Roy, Agostino Capponi, and Vineet Goyal (Columbia University) propose PREFER (Personalized Review Summarization with Online Preference Learning). The framework treats user preferences as unknown latent variables and uses an online learning algorithm to refine them as the user interacts with the summaries. For example, if a user consistently reads summaries highlighting battery life on electronics, PREFER will surface more battery-focused content in future summaries.

In a case study using the Amazon Reviews '23 dataset, controlled simulations showed that PREFER's online preference learning significantly improves alignment with target user interests compared to static baselines, without sacrificing summary quality (measured by faithfulness and coverage). The system does not require explicit user profiles or prior feedback—it learns implicitly from which summaries users engage with. This makes it practical for real-world e-commerce platforms where preferences are both diverse and dynamic. The paper also connects to game theory and optimization, suggesting broader applications in recommendation systems and adaptive interfaces.