Multi-Task and Multi-Scenario Algorithms for Recommendation Systems: Methods, Challenges, and Applications

The presentation introduces the application of multi‑task and multi‑scenario algorithms in recommendation systems, emphasizing challenges like conversion‑stage sparsity and the need for collaborative improvement across tasks and scenarios.

Huawei’s advertising ecosystem spans diverse media (own platforms and third‑party sites) and ad formats (APP, form, product), where user actions—from click to download and activation—create a long, sparse conversion chain.

For multi‑task modeling, the TAML (Task‑Adaptive Multi‑Level) framework is described, featuring multi‑level expert networks (shared, task‑specific, learner‑specific) and a distillation module that aggregates knowledge from multiple learners to enhance prediction robustness.

Multi‑scenario modeling is covered through CausalInt and DFFM approaches. These include a common‑feature extraction branch, a negative‑impact removal module using meta‑learning (MAML) and orthogonalization, and a scenario‑information migration module that employs causal intervention and dynamic weighting to transfer knowledge across scenes.

Experimental results show significant offline gains on public and Huawei private datasets compared with SOTA baselines, and online A/B tests demonstrate improvements in CVR, eCPM, and overall ad revenue across both own and third‑party media.

The Q&A section addresses practical concerns such as incremental training stability, additional learning signals for expert networks, and adaptive scenario planning strategies.