FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation

2026-05-20

AI Summary

This work addresses the challenge of short-lived live-streaming rooms in recommendation systems, where transient item lifecycles keep ID-based embeddings in a perpetual cold-start state, hindering effective generalization. To overcome this limitation, the authors propose LUCID, a novel framework that entirely eliminates item IDs in an industrial-scale system. Instead, LUCID leverages a cross-domain multimodal encoder to generate hierarchical discrete semantic codes at both room and segment levels, which are integrated into the ranking model via late fusion. The approach combines cross-domain joint training, online incremental learning, and a staged warm-up strategy to enhance recommendation performance. Empirical results demonstrate significant improvements: a 0.55% increase in high-quality watch time, a 2.05% boost in views for cold-start live rooms, and a 0.05% gain in user active duration.

Abstract

Modern recommender systems rely heavily on ID-based collaborative filtering: each item is represented by a unique ID embedding that accumulates collaborative signals from user interactions. Livestreaming recommendation, however, faces a unique challenge in this paradigm: a live room typically broadcasts for only tens of minutes, so its item ID remains poorly learned in a persistent cold-start state and ID-centric ranking models fail to generalize. We present FLUID, the first framework to fully retire the candidate-side item ID from a production-scale livestreaming ranker. FLUID couples a cross-domain multimodal encoder, jointly trained on short videos and livestreams to produce discrete hierarchical codes (LUCID), with a late-fusion, ID-free design that injects slice-level and room-level LUCID as independent tokens, stabilized by a staged warmup under online incremental training. Deployed on our industrial livestreaming recommenders with a cross-platform combined user base of over one billion globally, FLUID delivers significant online gains of +0.55% Quality Watch Duration, +2.05% Cold-Start Room Views, and +0.05% Active Hours.

Problem

• livestreaming recommendation
• cold-start
• ID-based collaborative filtering
• ephemeral items
• multimodal semantic representation

Innovation

• ID-free recommendation
• multimodal semantic codes
• livestreaming cold-start
• cross-domain encoder
• late-fusion architecture