Highlights of Recent Alibaba Advertising Research Papers Presented at WSDM 2022

WSDM (The International Conference on Web Search and Data Mining) is a top‑tier international conference in information retrieval and data mining, jointly organized by the SIGIR, SIGKDD, SIGMOD and SIGWEB committees. The 2022 edition will be held from February 21 to February 25 in the United States. This year the conference received 786 long‑paper submissions and accepted 159, resulting in an acceptance rate of approximately 20.23%.

Alibaba’s advertising technology team has four papers accepted at WSDM 2022. The authors will be invited to provide detailed analyses of their research ideas and technical achievements.

Leaving No One Behind: A Multi‑Scenario Multi‑Task Meta Learning Approach for Advertiser Modeling

Abstract: Advertisers play a crucial role on e‑commerce platforms such as Taobao and Amazon. While most academic work focuses on user‑side modeling (e.g., CTR prediction), advertiser‑side modeling—capturing diverse advertiser needs and campaign performance—has received far less attention. Advertiser modeling involves multiple tasks (e.g., spend, daily active users, clicks) across multiple marketing scenarios (search, display, live streaming). Directly modeling each scenario/task separately incurs high maintenance cost, suffers from data sparsity in small or new scenarios, and ignores complex inter‑scenario relationships that may vary with the task. To address these challenges, we propose M2M, a meta‑learning‑based multi‑task multi‑scenario modeling method that can predict multiple tasks across multiple scenarios simultaneously. M2M incorporates a meta‑learning unit to explicitly model scenario relationships, integrates this unit into both the attention layer (to capture dynamic scenario relations under different tasks) and the tower network (to strengthen scenario‑specific representations). Offline experiments on Alibaba’s industrial dataset and online A/B tests on the “Jiasu Bao” product demonstrate significant improvements in key business metrics such as daily active users and ARPU. The model has been fully deployed in Alibaba’s advertising ecosystem.

CAN: Feature Co‑Action Network for Click‑Through Rate Prediction

Abstract: Feature interaction is vital for click‑through rate (CTR) prediction. Deep neural networks (DNNs) can automatically learn implicit non‑linear interactions from sparse raw features, but they often fail to capture explicit relational information. Explicit interactions such as the Cartesian product of feature A and feature B (as used in Factorization Machines) improve performance but dramatically increase model parameters and data requirements. We propose the Feature Co‑Action Network (CAN), which approximates explicit feature interactions at a low cost. In CAN, given feature A and its related feature B, a Co‑Action Unit learns two parameter sets: (1) a high‑dimensional embedding for feature A, and (2) a multi‑layer perceptron (MLP) for feature B. Their interaction is obtained by combining the embedding and the MLP output. When the MLP depth is set to 1, CAN reduces to a standard FM. Experiments on industrial and public datasets show that CAN outperforms state‑of‑the‑art CTR models. Deployed in Alibaba’s display advertising system, CAN yields 12% and 8% improvements in CTR and RPM respectively.

Adaptive Unified Allocation Framework for Guaranteed Display Advertising

Abstract: Guaranteed Display (GD) contracts are widely used in e‑commerce marketing to guarantee a certain number of ad impressions for advertisers. Beyond maximizing contract fulfillment, user interest signals such as click‑through rate and conversion rate are essential for long‑term ROI. We design an Adaptive Unified Allocation Framework (AUAF) that jointly optimizes contract fulfillment and the match between ads and user interest while explicitly preventing over‑allocation of impressions. The model balances request‑level display volume with contract constraints. To handle billions of daily ad requests, we develop a parallel optimization algorithm based on parameter server (PS) that trains and updates the model within minutes, enabling synchronized offline training and online decision making. Offline and online experiments demonstrate that AUAF improves contract fulfillment and average CTR by over 10% without sacrificing fulfillment rates, delivering significant business value.

A Cooperative‑Competitive Multi‑Agent Framework for Auto‑bidding in Online Advertising

Paper download: https://arxiv.org/pdf/2106.06224.pdf

Abstract: Automated bidding is essential for advertisers to optimize performance in online advertising. Existing works largely treat auto‑bidding as a single‑agent problem, ignoring interactions among multiple advertisers. We study auto‑bidding from a distributed multi‑agent perspective and propose MAAB (Multi‑Agent Auto‑Bidding), a general framework for learning bidding strategies. MAAB introduces a temperature‑controlled reward allocation mechanism to create a mixed cooperative‑competitive relationship among agents, balancing individual advertiser utility with overall social welfare. To prevent sub‑optimal cooperative behavior that harms revenue, we add threshold agents that provide personalized bidding policies for each advertiser. For scalability, we employ a mean‑field approach that groups advertisers with similar goals into a single mean‑field agent, simplifying interactions and enabling efficient training. Experiments on industrial offline datasets and Alibaba’s advertising platform show that MAAB surpasses baseline algorithms in both advertiser welfare and platform profit.