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

Xiaohongshu, a rapidly growing UGC community in China, generates billions of content exposures daily, with a dual‑column feed and video immersion flow, making decentralized distribution a critical problem.

The core issues are twofold: the system must learn faster to capture long‑tail content and user interests, and it must learn better to ensure these signals are effectively propagated through recall, coarse ranking, fine ranking, and post‑ranking stages.

To accelerate learning, Xiaohongshu upgraded its recommendation pipeline from daily to minute‑level updates, introducing GPU‑heterogeneous training for early‑fine ranking models.

Link analysis revealed over‑reliance on note IDs, insufficient content‑signal utilization, and self‑reinforcing feedback loops, prompting solutions that enhance side‑information (sideinfo) usage, incorporate multimodal signals, and protect user interests.

Sideinfo decoupling separates sideinfo modeling from note ID embeddings, applying dedicated attention mechanisms and residual connections in both recall and ranking modules, and further integrates sideinfo into graph models with CF‑based edges, swing algorithm upgrades, and causal‑inference bias correction.

Multimodal signal fusion progresses from contrastive learning to the AlignRec framework, aligning text and image embeddings with item embeddings, and employing multimodal cross‑feature interactions in recall, coarse ranking, and fine ranking stages.

Interest exploration shifts from implicit multi‑vector methods to explicit global interest sets, selecting top‑k interests based on recent user behavior and predicting the next interest, while exploration‑exploitation (EE) adds Gaussian noise to recall vectors and uses evolution‑strategy‑driven regeneration of high‑potential items.

Interest protection introduces a white‑box multi‑queue framework for the intermediate ranking stages, allowing customized ordering and automatic cutoff optimization via evolution strategies.

Future work includes generative recommendation, interactive search with large models, multimodal user profiling, end‑to‑end multimodal‑behavior joint training, and full‑domain signal modeling across the entire app.

Overall, the presentation showcases Xiaohongshu's ongoing efforts to achieve decentralized distribution through algorithmic innovations and large‑model techniques.