Overview of Recent Alibaba Mama Research Papers Presented at KDD 2021 on Advertising and AI

ACM SIGKDD (the International Conference on Knowledge Discovery and Data Mining) is a top-tier conference in the data mining field, organized by the ACM SIGKDD committee and recommended as an A‑class conference by the China Computer Federation. Since its inception in 1995, the conference has been held 26 times, and the 2021 edition took place in Singapore from August 14‑18.

The 2021 KDD received 1,541 paper submissions, of which 238 were accepted (acceptance rate 15.44%).

Alibaba Mama’s technical team had six papers accepted, covering topics such as deep learning, bidding strategy recommendation, end‑to‑end mechanism optimization, and cooperative bidding games. The team also organized two workshops (DLP‑KDD and AdKDD) during the conference.

A Unified Solution to Constrained Bidding in Online Display Advertising

Abstract: In online display advertising, advertisers aim to maximize the value of won traffic under budget and KPI constraints. Existing work often focuses on a single demand and lacks generality. This paper formalizes various advertiser demands as a constrained bidding problem and derives a unified optimal bidding strategy. The optimal bid formula contains m parameters (one per constraint). Because the environment is dynamic, the paper proposes a reinforcement‑learning method that continuously adjusts the parameters to approximate the optimal solution. The approach, called USCB, achieves strong performance on industrial datasets and has been deployed in Alibaba Mama’s ad‑placement platform, improving both platform revenue and advertiser outcomes.

NeuralAuction: End‑to‑End Learning of Auction Mechanisms for E‑Commerce Advertising

Abstract: Traditional auction mechanisms (e.g., GSP, VCG) optimize a single objective and are sub‑optimal for multi‑stakeholder goals. This work proposes Deep Neural Auction (DNA), an end‑to‑end learnable auction mechanism that embeds the allocation process into a differentiable neural network using a relaxed sorting operator. The model also incorporates incentive‑compatible (truthful) constraints. Deployed in Taobao’s ad system, DNA outperforms classic mechanisms on both offline datasets and online A/B tests across multiple stakeholder metrics.

Real Negatives Matter: Continuous Training with Real Negatives for Delayed Feedback Modeling

Abstract: Conversion rate (CVR) prediction suffers from delayed feedback, where conversions may occur long after a click. Existing methods use short waiting windows or duplicate samples, leading to distribution shift and bias. This paper introduces Defer, which injects real negative samples into the training pipeline and uses importance sampling to correct distribution bias. Experiments on industrial datasets show >6% CVR lift, and Defer has been deployed in Alibaba’s display ad system.

We Know What You Want: An Advertising Strategy Recommender System for Online Advertising

Abstract: Advertisers are a key revenue source for e‑commerce platforms. To reduce trial‑and‑error in real‑time bidding, the authors built a recommendation system that models the advertiser’s strategy selection as a contextual bandit problem. A neural network predicts advertiser demand from historical behavior; simulated auctions derive optimal bidding strategies, and dropout‑based uncertainty estimates enable effective exploration, akin to Thompson sampling. Online evaluation shows increased advertiser adoption and platform revenue; offline simulations confirm the contextual bandit’s advantage.

Multi‑Agent Cooperative Bidding Games for Multi‑Objective Optimization in e‑Commercial Sponsored Search

Abstract: Real‑time bidding in e‑commerce search advertising involves millions of advertisers with diverse objectives. Existing single‑advertiser approaches assume competitors’ bids are static, which is unrealistic. This paper proposes a multi‑agent cooperative bidding framework with a global objective to align advertisers’ interests while constraining platform revenue to prevent collusion. The authors derive a functional form for the optimal bid and design a multi‑agent evolutionary strategy search to find optimal parameters. Extensive offline and online experiments on Taobao’s search ad platform demonstrate >5% improvement in both platform efficiency and individual advertiser goals.

Exploration in Online Advertising Systems with Deep Uncertainty‑Aware Learning

Abstract: State‑of‑the‑art online ad systems rely on deep CTR models but suffer from insufficient exploration. The authors propose Deep Uncertainty‑Aware Learning (DUAL), which augments a deep model with a Gaussian‑process‑based uncertainty estimator. DUAL can be added to existing models with minimal overhead. By combining DUAL’s uncertainty with bandit algorithms, the paper introduces a DUAL‑based ad‑placement strategy that improves long‑term utility. Experiments on public datasets and online A/B tests on Alibaba’s display ad platform show significant revenue gains.

All papers are available for download via the provided arXiv links.

In the past three years, Alibaba Mama’s team has published over 50 papers at top conferences. Notably, the Deep Interest Network (DIN) presented at KDD 2018 became a seminal work for CTR prediction and recommendation, and its code has been open‑sourced. The team continues to develop AI platforms such as XDL, MDL, and EULER, focusing on areas like intelligent matching, behavior prediction, mechanism design, smart bidding, smart creative, multi‑touch attribution, federated learning, and graph neural networks.