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

2026-05-23 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

FLUID is a novel framework addressing the persistent cold-start challenge in industrial-scale livestreaming recommendation. Live rooms are ephemeral, typically lasting only ~45 minutes, making traditional ID-based collaborative filtering ineffective. FLUID fully retires the candidate-side item ID. It integrates a cross-domain multimodal encoder, jointly trained on short videos and livestreams, producing discrete hierarchical semantic codes called LUCID. These slice- and room-level LUCID codes are injected as independent tokens into a late-fusion, ID-free ranker. A staged warmup stabilizes this transition under online incremental training. Deployed on platforms with over one billion users, FLUID achieved significant online gains. These included +0.55% Quality Watch Duration, +2.05% Cold-Start Room Views, and +0.05% Active Hours.

Key takeaway

For MLOps Engineers or AI Scientists building recommender systems for short-lived content like livestreams, you should consider fully retiring ephemeral item IDs. Implementing a cross-domain multimodal encoder that generates discrete semantic codes (LUCID) and integrating them via a late-fusion, ID-free ranker with a staged warmup can significantly boost engagement, cold-start exposure, and user retention, rather than merely supplementing existing ID-based systems.

Key insights

FLUID replaces ephemeral item IDs in livestreaming recommendation with cross-domain multimodal semantic codes to overcome cold-start limitations.

Principles

• Ephemeral item IDs in livestreaming cause persistent cold-start issues for ID-based collaborative filtering.
• Multimodal features are often underutilized when coexisting with dominant item IDs in ranking models.
• Cross-domain training enhances multimodal encoder generalizability for sparse, noisy live data.

Method

FLUID employs a cross-domain multimodal encoder (SigLIP2 ViT + Qwen3-Embedding) to generate 128-d slice embeddings, discretizes them into 4-level LUCID codes via RQ-KMeans, and injects slice- and room-level LUCID as independent tokens into a late-fusion, ID-free ranker with a staged warmup.

In practice

• Implement cross-domain training for multimodal encoders to improve performance on sparse, noisy datasets.
• Utilize discrete hierarchical codes (LUCID) to replace ephemeral item IDs in short-lived content recommendation.
• Employ a staged warmup procedure for stable transitions when replacing core features in online incremental training.

Topics

• Livestreaming Recommendation
• Multimodal Embeddings
• Cold Start Problem
• Semantic Codes
• Large Recommendation Models
• Cross-domain Learning

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• DiffCold: A Diffusion-based Generative Model for Cold-Start Item Recommendation — DiffCold employs conditional diffusion to unify warm and cold item representations, resolving the "seesaw dilemma" in recommendation systems.
• Guiding Federated Graph Recommendation with LLM-encoded knowledge — LLM-encoded semantic knowledge can guide federated graph recommendation, improving accuracy by selectively aggregating structural representations across clients.
• Inside ByteDance’s Monolith: The Engine Powering Smarter, Faster Content Feeds — Monolith optimizes real-time recommendation systems via collisionless embeddings and continuous online training.
• From Random IDs to Semantic IDs: Building a Generative Recommender from Scratch — Semantic IDs transform arbitrary item identifiers into meaningful, content-rich tokens for generative recommenders.
• PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation — PEARL uses contrastive learning to estimate unbiased percentile-based preferences, mitigating behavioral intensity bias in large-scale recommender systems.
• Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation — An asymmetric graph architecture can bridge the semantic-collaborative gap for cold-start recommendations by aligning intrinsic features with behavioral signals.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.