Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting

Live-streaming platforms like Twitch face unique challenges in delivering personalized content that balances immediate user engagement, long-term retention, and monetization. Traditional recommendation systems, often optimized for single objectives, struggle to handle the complex, multi-faceted nature of user behaviors in live environments. User actions such as watching, chatting, following, and spending occur at different times and frequencies, with high-value behaviors like subscriptions being particularly sparse and delayed. This creates a fundamental difficulty: models trained on immediate feedback may not capture the full spectrum of user preferences, especially for new or less active users.

To address these issues, the authors propose a comprehensive multi-objective ranking framework that combines several innovative components. Central to this approach is the delayed window mechanism, which extends feedback collection over a 14-day period, transforming sparse high-value actions into denser signals. This mechanism effectively mitigates the problem of behavior sparsity and delays, providing richer training data for the models.

Complementing this, the system employs a multi-model architecture that separates immediate signals (via Fresh Signal Models, FSM) from delayed signals (via Delayed Signal Models, DSM). During inference, predictions from both models are fused, ensuring responsiveness to recent user preferences while leveraging historical behavior. To further refine personalization, a Segment-Aware Targeting (VST) module adjusts ranking weights based on user lifecycle stages, such as new or highly engaged users. This addresses user bias and tailors recommendations to different segments.

At the core of the system is the Multi-gate Mixture-of-Experts (MMoE) architecture, which jointly models multiple deep engagement and monetization targets. By sharing expert networks and employing gating mechanisms, MMoE reduces model parameters by 41.9% compared to independent models, while maintaining high predictive accuracy. The entire system is optimized for low latency, with response times below 110ms, supporting large-scale deployment.

Extensive offline and online experiments on Twitch demonstrate the effectiveness of the proposed framework. Key metrics such as DAU increased by 0.09%, translating to millions of additional active days annually. High-value user actions, including new follows and capped ARPU, saw improvements of 0.27% and 0.56%, respectively. The system also generalized well to mobile live feeds, boosting positive user interactions by 1.12%. These results confirm that integrating delayed feedback, multi-model fusion, and segment-aware strategies can substantially enhance live-streaming recommendation performance.

Overall, this research offers a scalable, efficient solution for multi-objective recommendation in complex, real-time environments. It advances the state-of-the-art by explicitly modeling behavior delays and user segments, providing a foundation for future innovations in personalized content delivery across diverse platforms. The framework's flexibility and robustness make it a promising candidate for broader application in online entertainment, social media, and beyond. Future directions include incorporating multi-modal content features, reinforcement learning for dynamic weighting, and extending the approach to new domains, aiming to further push the boundaries of intelligent recommendation systems.