Overview of Alibaba Mama’s Recent Papers on Online Advertising and Recommendation Systems

Alibaba Mama’s technical team has had ten papers (eight long papers and two short papers) accepted at the 31st ACM International Conference on Information and Knowledge Management (CIKM 2022). The collection showcases the latest advances in online advertising algorithms and large‑scale recommendation techniques.

Adaptive Domain Interest Network for Multi‑domain Recommendation (ADIN) : Proposes a network that jointly models shared and domain‑specific patterns via shared and private subnetworks, scene‑aware batch normalization, and a self‑supervised training strategy, achieving a 1.8% revenue lift in Alibaba’s display ad system.

Approximate Nearest Neighbor Search under Neural Similarity Metric for Large‑Scale Recommendation : Extends ANN search to arbitrary neural similarity metrics using greedy walks on a similarity graph and a plug‑in adversarial training task, deployed for the Double‑11 shopping festival.

Control‑based Bidding for Mobile Livestreaming Ads with Exposure Guarantee : Formulates livestream ad bidding as an online integer programming problem, solves for optimal dual variables with a deep neural network, and introduces a real‑time control algorithm that satisfies exposure constraints.

Graph‑based Weakly Supervised Framework for Semantic Relevance Learning in E‑commerce : Builds a weak‑supervised contrastive learning framework that leverages heterogeneous user‑interaction graphs to generate semantic supervision and improves relevance ranking via hybrid fine‑tuning and transfer learning.

Hierarchically Constrained Adaptive Ad Exposure in Feeds (HCA2E) : Models dynamic ad placement as a dynamic knapsack problem with hierarchical constraints, delivering near‑optimal platform performance while respecting auction mechanisms.

KEEP: An Industrial Pretraining Framework for Online Recommendation via Knowledge Extraction and Plugging : Introduces a two‑stage pipeline that extracts knowledge from a super‑domain through supervised pre‑training and plugs it into downstream models, boosting CTR prediction in large‑scale industrial systems.

STARDOM: Semantic Aware Deep Hierarchical Forecasting Model for Search Traffic Prediction : Combines semantic embeddings and hierarchical calibration learning to improve search ad traffic forecasting across brands and categories.

Understanding the Overfitting Phenomenon of Deep Click‑Through Rate Models : Analyzes the “one‑epoch” overfitting behavior of deep CTR models, identifying fast convergence, large learning rates, and data sparsity as key factors.

AdaSparse: Learning Adaptively Sparse Structures for Multi‑Domain CTR Prediction : Learns per‑domain sparse sub‑networks with neuron‑level importance weights, reducing computation while enhancing cross‑domain generalization.

Visual Encoding and Debiasing for CTR Prediction : Proposes a contrastive learning based image representation framework and a debiasing network to improve CTR prediction accuracy and fairness in image‑search advertising.

All papers have been deployed in Alibaba’s advertising platforms, delivering measurable revenue and performance improvements.