FLUID: ID-Free Framework for Livestreaming Recommendation

Researchers have introduced a novel framework named FLUID, as outlined in arXiv paper 2605.21832, which eliminates the need for item IDs in livestreaming recommendation systems. Conventional collaborative filtering that relies on IDs struggles with live rooms, which typically last only a few minutes and are perpetually in a cold-start phase. FLUID employs a cross-domain multimodal encoder, trained on both short videos and livestreams, to generate discrete hierarchical codes known as LUCID. This approach features an ID-free design with late fusion, incorporating slice-level and room-level LUCID as distinct tokens, supported by staged warmup during online incremental training. FLUID has been implemented in large-scale livestreaming recommendation systems.