DiffuRetrieval: A Chain-of-Thought Enhanced Diffusion Retrieval in Sponsored Search

Embedding-based Retrieval (EBR) system is a fundamental component that supplies candidates for downstream ranking mechanisms in the sponsored search system. To enhance search experience and ensure effective retrieval, EBR usually accounts for various objectives including the semantic relevance and personalization of search results. However, traditional multi-task EBR models ignore the intrinsic progressive relationship between relevant and personalized candidates during a search. Recognizing this gap, we make the very first attempt to utilize the representation generation capabilities of Diffusion Models in EBR. In this paper, we present a novel model DiffuRetrieval to address the progressive objectives for high-quality item retrieval. In forward process, DiffuRetrieval incrementally corrupts item representations through controlled noise injection. Conversely, in reverse process, we refine the representations based on query information in a chain-of-thought manner, initially establishing coarse-grained relevance and progressively moving towards fine-grained personalization. Online A/B tests on Meituan sponsored search platform demonstrate that our approach markedly surpasses the baselines, delivering substantial improvements in revenue, relevance and personalization.