Cross‑Scene Intelligent Advertising on Xiaohongshu: Algorithms for Keyword Selection, Targeting, Budget Allocation and Multi‑Constrained Bidding

Xiaohongshu, a lifestyle content and consumer‑decision platform, offers multiple advertising scenes (e.g., Discover feed and Search feed). To simplify cross‑scene campaign management, the company built the “All‑Site Smart投 (全站智投)” system, which provides a one‑stop, AI‑driven solution for intelligent keyword extraction, targeting, budget distribution, and bidding.

Intelligent Keyword Extraction : The system uses a two‑stage “recall + selection” pipeline. In the recall stage, content understanding (core‑word extraction, semantic parsing, multimodal representation) and user‑behavior modeling (graph neural networks) jointly generate candidate keywords. In the selection stage, the candidates are treated as an Exploit‑and‑Explore (E&E) problem, where each keyword is repeatedly sampled, evaluated with online feedback, and its selection probability is updated to maximize long‑term ROI.

Handling Sparse Ad Formats : For ad formats with limited textual information (e.g., forms, product cards), OCR, image classification, and query rewriting are applied to enrich visual material. Contrastive learning and graph‑based relationships (same‑advertiser, co‑occurrence) are incorporated to improve modeling of these sparse samples.

Intelligent Targeting : Instead of requiring advertisers to specify explicit audience segments, the system performs Learning‑to‑Rank (LTR) based recall, aligning the recalled traffic with the advertiser’s implicit goals. A pointwise LTR model predicts a delivery probability (pSend) using a dual‑tower architecture, which guides the downstream ranking.

Budget Allocation Across Scenes : Two strategies are discussed. (1) A unified bidding module that bids at the traffic‑level using scene‑specific value estimates (pValue). (2) Separate budget allocation per scene followed by independent intelligent bidding. Xiaohongshu adopts the second approach for better system decoupling and robustness. A heuristic dynamic adjustment algorithm reallocates budgets based on consumption rate, cost, and constraint violations.

Multi‑Constrained Bidding (MCB) : Advertisers provide overall budget, target cost, and optimization objective. The platform formulates bidding as a linear‑programming problem with multiple constraints, solves it via a primal‑dual method in the dual space, and derives a closed‑form optimal bid that includes a tunable hyper‑parameter. Each virtual plan runs an online MCB model that continuously updates bid parameters using historical performance (cost‑achievement, budget‑spend rate) and predicted auction conditions.

The system has been deployed across Xiaohongshu’s various ad scenes, and future work includes extending the cross‑scene unified bidding architecture and further upgrading the intelligent components.