跳转至

简体中文 · English

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

Core mechanism reproduction. The defining operator runs locally, without private data, large model weights or dedicated serving. Paper results are separate.

Paper information

Field Details
Paper arXiv 2608.04455
Institution Twitch
First public release 2026-08-05 (arXiv v1)
Author code No official author code in the paper when checked on 2026-08-09.
Local adapter or method key twitch-mor
Local reproduction code src/auto_research/reproductions/twitch_mor/

Original paper summary

Background and main changes

Topic: Multi-objective livestream ranking. Viewing, interaction, follows and payment arrive at different delays and vary across lifecycle segments. The paper combines immediate and delayed models, segment-specific objective weights and MMoE with shared experts and objective-specific gates.

flowchart LR
    N0["Immediate behavior"] --> N1
    N1["Delay window"] --> N2
    N2["Lifecycle segment"] --> N3
    N3["MMoE multitask gates"] --> N4
    N4["Fused ranking"]

Key figure from the paper

twitch-mor Figure 1

Figure 1: Training workflow and key optimizations. From the paper; copyright belongs to the authors. Click for the source.

Core equations

\(y_k=\sum_e g_{k,e}(x)f_e(x),\qquad \mathcal L=\sum_k\lambda_k(s)\mathcal L_k\)

Offline and online results

Production DAV rises +0.09% and capped ARPU for highly engaged users +0.56%. MMoE also yields DAV +0.08%, new follows +0.27% and positive engagement in the mobile live feed +1.12%.

Local reproduction

Create five immediate/delayed objectives from MovieLens-1M. Compare a single-objective DNN with a three-expert lifecycle MMoE.

auto-research reproduce --paper twitch-mor --dataset-dir data --seed 42

See stable metrics. No checkpoint is committed.

Comparison: Remove the paper-specific mechanism while matching split and budget; the experimental arm uses twitch-mor. Relative change is in public-seed42.json; cross-paper percentages are not comparable.

Scope

  • Local results verify execution and direction, not paper scale.
  • Private features, traffic and serving are unavailable; paper online figures are references only.
  • Evolve-compatible architectures are registered as candidates; serving-only methods are standalone runnable adapters, not trainable genomes.