Decentralized Distribution in Xiaohongshu: Strengthening Sideinfo, Multimodal Fusion, and Interest Exploration

Xiaohongshu, a rapidly growing UGC platform, faces a centralized distribution problem where popular content dominates exposure, suppressing long‑tail creators and niche user interests. The article defines the decentralization challenge and proposes solutions from both content and user sides.

Content‑side improvements include strengthening sideinfo usage by decoupling its modeling, applying residual attention, and extending the technique to ranking modules. Graph models are employed to fuse sideinfo with collaborative‑filtering edges, mitigate popularity bias via swing algorithms, and use causal inference to adjust edge weights. Multimodal signals are incorporated through contrastive learning, the AlignRec framework, and feature crossing in recall and ranking stages, achieving better long‑tail content coverage.

User‑side strategies focus on interest exploration and protection. Explicit multi‑interest modeling builds a global interest set and selects top‑k interests for each user, while evolutionary‑strategy‑based exploration adds Gaussian noise to recall vectors and re‑injects high‑potential items. A white‑box multi‑queue framework preserves explored interests across intermediate ranking stages, and large‑model‑driven potential interest mining uses prompts and offline inference to boost downstream ranking.

The article concludes with future directions such as generative recommendation, interactive search with large models, multimodal user profiling, end‑to‑end multimodal‑behavior joint training, and global signal modeling across the entire app ecosystem.