Alibaba Mama Team Papers Accepted at KDD 2024

Recently, the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) announced its paper acceptance results, and Alibaba Mama's technical team had four papers accepted.

KDD 2024, a CCF‑recommended A‑class international conference and a top venue in data mining, will be held in Barcelona, Spain from August 25‑29. The research track received 2,046 submissions with an acceptance rate of about 20%.

The following are the abstracts of the four accepted papers.

▐ AIGB: Generative Auto‑bidding via Diffusion Modeling

Authors: Jiayan Guo, Yusen Huo, Zhilin Zhang, Tianyu Wang, Chuan Yu, Jian Xu, Yan Zhang, Bo Zheng

Link: https://arxiv.org/abs/2405.16141

Abstract: Automatic bidding is crucial for online advertising efficiency. Existing reinforcement‑learning (RL) approaches suffer from the Markov assumption and error propagation in long‑horizon decisions. This paper proposes a generative bidding paradigm, AI‑Generated Bidding (AIGB), which employs a conditional diffusion model (DiffBid) to directly model the relationship between returns and ad‑placement trajectories, avoiding error propagation. DiffBid can generate trajectories that respect constraints while maximizing objectives. Extensive experiments on real datasets and online A/B tests on Alibaba’s advertising platform show that DiffBid improves GMV by 2.81% and ROI by 3.36% over traditional RL methods.

▐ Truthful Bandit Mechanisms for Repeated Two‑stage Ad Auctions

Authors: Haoming Li, Yumou Liu, Zhenzhe Zheng, Zhilin Zhang, Jian Xu, Fan Wu

Abstract: Online advertising often uses a two‑stage auction: a first stage selects a promising subset of ads, and a second stage runs a refined auction on that subset. Designing a truthful, individually rational mechanism for repeated two‑stage auctions is challenging. We formulate the problem as a combinatorial bandit with a generic reward function and propose two mechanisms—an “explore‑then‑commit” scheme and a refined version with lower regret—both satisfying ex‑post truthfulness and ex‑post IR, while providing regret guarantees.

▐ Bi‑Objective Contract Allocation for Guaranteed Delivery Advertising

Authors: Yan Li, Yundu Huang, Wuyang Mao, Rongfu Ye, Xiang He, Zhonglin Zu, Shaowei Cai

Abstract: Contract advertising involves an order‑selling phase and a delivery phase. Existing works treat them separately, ignoring delivery constraints during allocation. We propose a bi‑objective allocation method that maximizes new order sales while balancing inventory distribution. A tailored local‑search algorithm efficiently handles the high‑dimensional, heavily constrained problem, outperforming multi‑objective evolutionary algorithms and commercial solvers such as Gurobi.

▐ An Efficient Local Search Algorithm for Guaranteed Delivery Advertising Contract Allocation

Authors: Xiang He, Wuyang Mao, Zhenghang Xu, Yuanze Gu, Yundu Huang, Zhonglin Zu, Liang Wang, Mengyu Zhao, Mengchuan Zou

Abstract: Large‑scale contract advertising faces non‑convex multiple linear constraints (e.g., media‑preference requirements). Traditional solvers struggle with these constraints in real‑time. We present a two‑phase local‑search framework with four novel operators designed for non‑linear constraints. Experiments demonstrate that our algorithm delivers high‑quality solutions within business time limits and can be adapted to other scenarios with similar constraints.