Decoding Xiaohongshu’s Recommendation System: How Ordinary Users Gain Visibility

The article explores the recommendation system of Xiaohongshu, a popular Chinese lifestyle platform, focusing on why ordinary users can easily become visible and how the platform balances user experience, community interaction, and content diversity.

It describes the platform’s shift from a simple “grass‑planting” tool to a comprehensive “life encyclopedia,” emphasizing a decentralized distribution strategy that prioritizes high‑quality user‑generated content (UGC) and multi‑modal data (text, images, video, live streams).

Key technical components include a large‑scale multimodal pre‑training model (similar to CLIP) that converts notes into vector representations, enabling fine‑grained content clustering and dynamic tag generation. Various backbone architectures (BERT, RoBERTa, ResNet, Swin‑T, ViT) are employed to support downstream tasks.

The cold‑start pipeline consists of four steps: (1) content information extraction using NLP, CV, and multimodal fusion; (2) seed‑user selection via dual‑tower models and graph neural networks, with Bayesian optimization for boost coefficients; (3) look‑alike user expansion based on behavior feedback, improving click‑through rate by ~7%; and (4) continuous model updates with minute‑level iteration for recall, coarse‑ranking, and fine‑ranking stages.

To ensure recommendation diversity, Xiaohongshu adopts exploitation‑exploration strategies, leveraging deterministic point processes (DPP) and multi‑granular sampling (MGS) to disperse similar items and capture long‑tail interests. The Sliding Spectrum Decomposition (SSD) model transforms per‑item scoring into a session‑level optimization, while the CB2CF method replaces traditional collaborative filtering with content‑based similarity for better long‑tail handling.

The article also discusses the integration of large language models (LLMs) into the recommendation pipeline, noting challenges such as latency, resource consumption, and explainability. Xiaohongshu prefers augmenting existing systems with LLMs (e.g., as feature encoders or control modules) rather than replacing the entire stack.

Overall, the piece provides a comprehensive view of how advanced AI techniques—multimodal embeddings, cold‑start pipelines, diversity models, and LLM augmentation—enable Xiaohongshu to deliver personalized, equitable content discovery for ordinary users.