UniGD Unifies Generative and Discriminative Retrieval for Ads.

This paper introduces UniGD, a Unified Generative-Discriminative framework designed to revolutionize industrial search advertising. Current generative retrieval (GR) systems often separate the generative likelihood objective from query-ad relevance discrimination, leading to suboptimal effectiveness and higher serving costs. UniGD addresses this by integrating both retrieval and relevance scoring within a single, cohesive model. To manage the potential gradient interference during joint optimization, UniGD incorporates Conflict-Aware Gradient Enhancement (CAGE), which adaptively coordinates the two objectives. Furthermore, it features a Codebook-Anchored Representation Module (CAM) that grounds item representations in frozen hierarchical codebooks derived from multimodal pretrained models, enriching them with robust semantic priors. For diverse ad types like short-video, product, and live-stream ads, UniGD employs Heterogeneous Ad-material Modeling (HAM). This component captures cross-type semantic commonality through a shared backbone while preserving type-specific modeling capabilities. Online A/B tests on a major search advertising platform demonstrated significant improvements, including a 5.78% increase in ad revenue, a 33% reduction in inference latency, and enhanced discriminative relevance estimation.