RCLRec: Reverse Curriculum Learning for Modeling Sparse Conversions in Generative Recommendation

A research team led by Yulei Huang has introduced RCLRec, a novel framework designed to solve a core problem in AI-driven recommendation systems: sparse conversion data. While modern generative recommendation (GR) models unify user behaviors into token sequences, they still struggle to effectively model rare but critical actions like purchases or sign-ups. RCLRec innovates by applying reverse curriculum learning. For a target conversion, it works backward through a user's history to construct a short, focused 'curriculum' of the most relevant preceding items. This curated sequence is fed to the model as a prefix, providing instance-specific intermediate supervision that directly trains the model on the user's decision pathway, not just their entire history.

This technical approach is more than theoretical. The framework includes a curriculum quality-aware loss function to ensure the selected historical snippets are genuinely informative for predicting the final conversion. The results are compelling. In offline dataset evaluations and, more importantly, a live online A/B test, RCLRec demonstrated superior performance. The deployed model delivered a significant lift in key business metrics, achieving a +2.09% increase in advertising revenue and a +1.86% rise in orders. This proves the method's effectiveness at turning sparse behavioral signals into concrete commercial outcomes.

The success of RCLRec highlights a shift from modeling broad user histories to intelligently pinpointing pivotal decision moments. By making AI recommendations more interpretable and focused on conversion causality, it provides a scalable blueprint for platforms where monetization depends on accurately predicting and influencing these high-value, infrequent user actions.