ResearchPod Summary
E-commerce platforms struggle with the 'cold-start' problem, where new users have insufficient interaction history to accurately predict their Lifetime Value (LTV) and Conversion Rate (CVR). Existing approaches, such as LLM-based semantic augmentation and Learning Using Privileged Information (LUPI), suffer from noisy, unstructured rationales and brittle distillation processes, respectively. This paper introduces SemRaD, a framework designed to bridge the information gap between sparse online signals and privileged post-conversion data.
SemRaD employs a two-part strategy to enhance prediction accuracy. First, it implements a Structured Semantic Reasoning Pipeline that replaces free-form LLM rationales with a structured schema. This schema is developed through a 'discover-curate-audit' workflow, ensuring that the semantic profiles generated for users are consistent, interpretable, and task-relevant. These profiles are then processed by a Semantic-Gated Encoder, which dynamically weights the most informative dimensions for each user.
Second, the authors propose a Hindsight-Aware Distillation Network. During training, a teacher model accesses both pre- and post-conversion logs, which are fused into a 'Hindsight Distillation Target.' This target reconciles potential contradictions between early behavior and final outcomes, providing the student model with more consistent supervision. Distillation experts are used to handle the heterogeneity of the information gap across different user segments.
SemRaD demonstrates significant performance gains on large-scale industrial datasets, achieving a +1.9% lift in LTV (Gini) and +1.0% in CVR (AUROC) compared to a production-grade baseline. A four-week online A/B test at Keeta confirmed these improvements, showing a +1.0% increase in LTV and +0.43% in CVR. Furthermore, the model exhibits high data efficiency, matching the production system's LTV performance while using only 9% of the training data.
By formalizing how LLMs contribute to predictive modeling through structured schemas and hindsight distillation, SemRaD provides a robust, production-ready solution for cold-start scenarios. The framework's ability to maintain high performance with significantly less data makes it a powerful tool for platforms looking to optimize marketing spend and user retention without relying on massive, long-term interaction histories.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.