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

This document introduces the use of multi‑task and multi‑scenario learning techniques in large‑scale recommendation systems, particularly within Huawei's advertising platforms.

It first outlines the background and challenges of multi‑task and multi‑scenario recommendation, emphasizing the sparsity of conversion samples across a cascade of user actions (click, download, activation, etc.) and the need for joint modeling to alleviate this sparsity.

The authors describe a taxonomy of algorithms. For multi‑task learning, they discuss task‑relationship dimensions (parallel, cascade, primary‑plus‑auxiliary) and methodological dimensions (hard/soft parameter sharing, task‑level optimization, training mechanisms such as reinforcement learning). The TAML framework is highlighted as a multi‑task solution that combines hierarchical expert networks with a distillation‑based multi‑learner module to improve robustness and address sparse conversion signals.

For multi‑scenario learning, the paper classifies approaches by model structure (multi‑tower vs. single‑tower) and method (parameter sharing, negative transfer mitigation, dynamic weighting). Two representative works are presented: CausalInt, which uses causal graphs and intervention‑style training to handle negative information transfer, and DFFM, a single‑tower model that incorporates dynamic weight networks for scene‑aware feature interaction (DFFI) and user‑behavior modeling (DFUB) using multi‑head attention.

The DFFM architecture consists of a scene‑aware feature interaction module that transforms raw feature vectors with scene embeddings via dynamic weight networks, and a user‑behavior module that enriches multi‑head attention with scene‑conditioned query/key/value matrices. The outputs of DFFI and DFUB are fused by a classifier (e.g., MLP) for final prediction.

Extensive offline experiments on public benchmarks and Huawei’s private data demonstrate consistent gains over state‑of‑the‑art baselines. Online A/B tests show improvements in click‑through rate, conversion rate, and eCPM across both owned and third‑party media placements.

The article concludes with a brief Q&A covering practical issues such as incremental training stability, additional learning signals for multi‑task experts, and strategies for handling dominant scenes in multi‑scenario training.