UniGD framework unifies generative-discriminative models for industrial search advertising

Researchers have developed UniGD, a novel framework that unifies generative and discriminative models for industrial search advertising. This approach aims to overcome the limitations of current systems that cascade these models separately, leading to suboptimal performance and increased costs. UniGD integrates retrieval and relevance scoring into a single model, employing techniques like Conflict-Aware Gradient Enhancement (CAGE) to manage conflicting objectives and a Codebook-Anchored Representation Module (CAM) for richer semantic understanding. Online A/B tests on Kuaishou's platform demonstrated a 5.78% increase in ad revenue and a 33% reduction in inference latency, alongside improvements on benchmark datasets like NQ320K and MS300K. AI

IMPACT This unified framework could significantly improve efficiency and effectiveness in industrial search advertising systems.

RANK_REASON The cluster contains a research paper detailing a new framework for generative retrieval. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Codebook-Anchored Representation Module
• Conflict-Aware Gradient Enhancement
• Generative Retrieval
• Heterogeneous Ad-material Modeling
• Kuaishou
• MS300K
• NQ320K
• UniGD

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →