New multi-objective live-stream recommender cuts parameters 41.9% and lifts engagement

A team of researchers (Xiaoyi Gu, Julia Tavares, Eder Santana, and colleagues) has published a paper detailing a new recommender system architecture designed specifically for live-streaming platforms. Unlike e-commerce, where user actions follow linear sequences, live-streaming viewers engage in concurrent behaviors—watching, chatting, following, and spending—each with varying delays. The system tackles this by introducing a delayed window approach that extends feedback collection beyond immediate responses, combined with a multi-model architecture that separately captures fresh and delayed signals.

The architecture also integrates segment-aware targeting, optimizing ranking scores differently across user lifecycle stages, and employs Multi-gate Mixture-of-Experts (MMoE) to jointly model correlated targets. This MMoE integration cuts model parameters by 41.9% compared to independent models. In online A/B testing, the system delivered a +0.09% increase in Daily Active Viewers (DAV), translating to millions of additional annual active viewer days, plus a +0.56% lift in capped ARPU from highly engaged viewers. Segment-aware targeting added +0.15% DAV for newer and less engaged users, while MMoE added +0.08% overall DAV and +0.27% new follows. The system was also validated on Twitch's mobile live feed, achieving a +1.12% increase in positive user-channel interactions. The paper will be presented at the Industry Track of RecSys 2026.