SSRLive uses dynamic semantic IDs to boost live streaming recommendations by 3.38% watch time

SSRLive addresses two core limitations in live streaming recommendation: static semantic IDs (SIDs) fail to capture fast-changing room content, and generative pipelines typically ignore user-streamer interaction signals critical for modeling intent. The framework unifies a generative module (encoder-decoder producing both static and dynamic SIDs using multimodal data) with a discriminative module that enriches task-specific representations with user features and interaction signals, enabling multi-task predictions (watch time, GMV, followers, interactions).

Real-world A/B tests on a major live platform (likely Tencent) yielded significant lifts: watch time +3.38%, GMV +0.72%, follower growth +3.12%, and interaction volume +2.92%. SSRLive is now fully deployed, serving hundreds of millions of active users. The paper demonstrates that dynamic SIDs plus interaction-aware discriminative modeling can materially improve business metrics without requiring massive FLOP increases, making it a practical solution for high-throughput live streaming environments.