PRECTR-V2:Unified Relevance-CTR Framework with Cross-User Preference Mining, Exposure Bias Correction, and LLM-Distilled Encoder Optimization
New framework tackles search's biggest problems: cold-start users, exposure bias, and frozen BERT bottlenecks.
A research team has introduced PRECTR-V2, a significant upgrade to their unified framework for coordinating search relevance matching and click-through rate (CTR) prediction. The new version directly addresses three persistent challenges in large-scale search systems: effectively modeling preferences for low-activity and new users, correcting the exposure bias caused by training only on high-relevance data, and overcoming the limitations of a frozen BERT encoder that prevents joint optimization with downstream CTR tasks. By solving these, PRECTR-V2 aims to create more consistent and effective ranking models that better discover user interests while enhancing platform revenue.
Technically, PRECTR-V2 employs a multi-pronged approach. It mines global relevance preferences under specific queries to personalize results for cold-start users. To correct exposure bias, it constructs hard negative samples through embedding noise injection and optimizes their ranking against positives. Most notably, it replaces the traditional frozen BERT module with a pretrained, lightweight transformer encoder. This new encoder is distilled from a large language model (LLM) and fine-tuned on text relevance classification, enabling it to be jointly optimized with the CTR prediction task, moving beyond the restrictive Emb+MLP paradigm and aligning representation learning with final ranking objectives.
- Solves cold-start user modeling by mining global relevance preferences under specific queries, personalizing results despite sparse behavior.
- Corrects exposure bias by constructing hard negative samples via embedding noise injection and optimizing with pairwise loss.
- Replaces frozen BERT with a lightweight, LLM-distilled transformer encoder, enabling joint optimization and breaking the Emb+MLP bottleneck.
Why It Matters
This could lead to more accurate, personalized search results for all users, especially new ones, while making ranking models more robust and efficient.