Multi‑Task and Multi‑Scenario Algorithms for Recommendation Systems – A Huawei Case Study

The presentation introduces the use of multi‑task and multi‑scenario algorithms in recommendation systems, focusing on Huawei's advertising platforms. It first outlines the background and challenges of sparse conversion data across diverse media such as Huawei Video, Browser, Music, and third‑party sources.

It then classifies algorithms into two dimensions: task‑relationship (parallel, cascade, primary‑plus‑auxiliary) and method (hard/soft parameter sharing, task optimization, training mechanisms). Representative works include the TAML multi‑task model and the CausalInt multi‑scenario framework.

The TAML architecture consists of a multi‑level task‑adaptive representation extractor and a multi‑learner distillation module, designed to alleviate sample sparsity and improve robustness. Detailed modules include shared experts, task‑specific experts, and learner‑level experts, whose outputs are fused by a gating network.

For multi‑scenario modeling, the CausalInt method uses causal intervention to separate scene‑invariant and scene‑specific features, addressing negative transfer and cold‑start issues. The DFFM model (with DFFI and DFUB modules) further incorporates dynamic weight networks for scene‑aware feature interaction and user‑behavior modeling, employing multi‑head attention that integrates target items and scene embeddings.

Extensive offline AB tests on public and Huawei private datasets, as well as online experiments on click‑through rate, conversion rate, and eCPM, demonstrate significant performance gains over single‑task or single‑scene baselines. Ablation studies highlight the contributions of each module.

The session concludes with a Q&A covering offline vs. incremental training, learning signals for different experts, and adaptive scheduling across dominant scenes.