Meta‑Network Based Multi‑Scenario Multi‑Task Model (M2M) for Alibaba Advertising Merchants

The paper addresses merchant understanding and modeling for Alibaba’s advertising ecosystem, a key component of the Alibaba Mama customer growth platform. Unlike user‑side modeling, merchant modeling involves diverse and sparse data, multiple business goals, and a variety of advertising scenarios (search, feed, brand search, live, interactive, etc.).

To meet these challenges, the authors propose a Meta‑Network based Multi‑Scenario Multi‑Task model (M2M) that can predict multiple targets across multiple scenarios in an end‑to‑end fashion. The model was presented at WSDM 2022.

Problem Definition : Given scene attributes, merchant profile, multi‑category behavior sequences, and multi‑category effect sequences, the goal is to predict, for the next t days, a set of tasks (e.g., shop spend, clicks, active days) under various advertising scenarios (e.g., Direct Traffic, Super Recommendation, Fast Push). The learning objective is to learn a function f that maps the input features to the future task values.

Algorithm Modeling : The architecture consists of two major parts:

Backbone Network : extracts embedded representations for experts, tasks, and scenes. Expert embeddings are obtained via a Transformer over dense and sequential features.

Meta‑Learning Network : includes three sub‑modules—Meta Unit (explicit scene modeling), Attention Meta‑Network (captures dynamic task‑scene relationships), and Tower Meta‑Network (enhances scene‑specific representations). The Meta Unit generates scene‑specific weight matrices and bias vectors that are fed as dynamic parameters to the downstream attention and tower networks.

The attention meta‑network computes attention scores by incorporating scene embeddings, allowing the model to adapt attention weights per scenario. The tower meta‑network then produces final task‑specific predictions.

Model Training : All tasks are regression problems with approximately Poisson‑distributed targets, so a Poisson loss is used. The overall loss combines weighted task losses with L2 regularization on both the meta‑network and the underlying multi‑task layers.

Experiments :

Offline experiments use SMAPE (Symmetric Mean Absolute Percentage Error) and NMAE (Normalized MAE) to evaluate performance across three scenarios (A, B, C) and tasks (clicks, spend, active days). The M2M model consistently outperforms baselines (MMoE, CGC, Cross‑Stitch), achieving improvements up to +26.7% in SMAPE and +24.1% in NMAE, especially in sparse‑data scenarios.

Online experiments on the “Accelerate Treasure” product demonstrate business impact: the upgraded model yields +2.59% increase in merchant activity rate and +2.09% uplift in per‑user ARPU.

Conclusion & Outlook : The M2M framework effectively handles multi‑scenario, multi‑task prediction with strong generalization, particularly in data‑sparse settings. Future work includes extending to classification tasks, improving universal applicability, and integrating optimization‑based meta‑learning approaches.

References :

[1] Ma J, Zhao Z, Yi X, et al. Modeling task relationships in multi‑task learning with multi‑gate mixture‑of‑experts. KDD 2018. [2] Tang H, Liu J, Zhao M, et al. Progressive layered extraction (PLE): A novel multi‑task learning model for personalized recommendations. RecSys 2020. [3] Zhao J, Du B, Sun L, et al. Multiple relational attention network for multi‑task learning. KDD 2019.