Reminder: ASA's marketing section's online seminar talk is tomorrow at 6PM Eastern Time.
Speaker: Yuyan Wang, Graduate School of Business, Stanford University
Title: Beyond Black-Box: Structuring Landing Page Recommender Systems Using Predicted Intents
Zoom link: usc.zoom.us/j/95561331935
Time & Date: 6pm-7pm Eastern Time, 11th November
Abstract: Modern recommender systems rely on black-box machine learning models to predict consumer choices. However, because these models do not explicitly represent the underlying data-generating process (DGP), they often struggle to generalize beyond observed data. A growing body of work advocates for incorporating consumer intent into personalization systems to improve generalization. Yet in the context of landing page recommendations - the most common and challenging personalization setting - a list of recommendations must be generated immediately when a consumer enters the platform, before any explicit intent signal is available. We propose the Intent-Structured Landing-Page Recommender System (ISRec), a principled framework that incorporates intent-based structure into multi-stage landing page recommender systems without requiring explicit consumer input and while satisfying industrial latency constraints. ISRec defines intent as a consumer's dynamic receptiveness to subsets of content, allowing intent labels to be inferred directly from observed behavior. It consists of three stages: intent prediction, intent-aware reward modeling, and intent-aware whole-page optimization, serving as the optimal greedy solution to the original NP-hard intent-aware recommendation problem with provable regret-bound guarantees. We evaluate ISRec on YouTube, the world's largest video recommendation platform, and find that it significantly improves daily active users (DAU) by 0.05% and overall user enjoyment by 0.09%, among the largest business metric gains observed in recent YouTube experiments, corresponding to an estimated $32.5 million in annual ads revenue. Our findings provide empirical evidence that even without knowing the true DGP, enforcing a partial structure aligned with it can help the model generalize. Managerially, ISRec offers a principled and generalizable framework for integrating behavioral and structural insights into real-time industrial personalization systems, paving the way for human-in-the-loop AI design.
------------------------------
Gourab Mukherjee
Associate Professor
University of Southern California
------------------------------