SSRLive: Dynamic semantic IDs for live-stream recommendation¶
Core public-data mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Taobao & Tmall Group, Alibaba. |
| First public release | 2026-06-05 (arXiv v1) |
| Author code | No public author code found as of 2026-09-02. |
| Local adapter or method key | ssrlive |
| Local reproduction code | src/auto_research/reproductions/ssrlive/ |
Paper method and evidence¶
Static SIDs encode creator content; dynamic SIDs capture current stream state. Fuse user–creator interactions for viewing, transaction and engagement tasks.
flowchart LR
A[Creator history and live state] --> B[Static and dynamic semantic IDs]
B[Static and dynamic semantic IDs] --> C[Fuse interaction features]
C[Fuse interaction features] --> D[Predict multiple outcomes]Original paper figure¶
Figure shows static and dynamic live semantic features. From the original paper; copyright remains with the authors.
Core equation¶
\[
h_i=\operatorname{Fuse}(e^{static}_{sid},e^{dynamic}_{sid},e_{u\leftrightarrow s}),\qquad \hat y_t=f_t(h_i).
\]
Reported results¶
Taobao Live A/B: watch time +3.38%, GMV +0.72%, follows +3.12%, engaged users +2.92%; optimized latency only 1.33% above DLRM.
Local reproduction¶
MovieLens-1M full catalog: NDCG@10 0.05401→0.05351 (-0.92%). Three-seed metrics, mean, SD, and 95% CI. No private logs, production models, bucketing or serving. Public proxies cannot establish online lift.