1 views collected around this technical thread.
AIGB, introduced by Alibaba Mama in 2023, reframes large‑scale ad‑auction auto‑bidding as a generative sequence task using diffusion models, achieving up to 5 % GMV gains, improved stability and interpretability, and is now commercialized, open‑sourced, and featured in a NeurIPS‑endorsed competition.
In real‑time bidding advertising, a multi‑path recall framework quickly filters billions of ads using parallel non‑personalized and personalized strategies—such as hot‑item rules, collaborative‑filtering, skip‑gram vectors, and GraphSAGE embeddings—while respecting targeting constraints, before a ranking stage optimizes eCPM, with effectiveness measured offline and online and future extensions planned with large language models.
The Mixed‑Modal Expert Model combines ID features with image and text embeddings through optimized representations and conditional output fusion, dramatically improving advertising recall—especially for long‑tail items—and delivering measurable gains in click‑recall, revenue, CTR, and page views in large‑scale online tests.
The advertising engine uses a five‑stage funnel—retrieval, recall, coarse ranking, fine ranking, and re‑ranking—each optimized with specialized indexes, multi‑channel recall, multi‑objective twin‑tower models, deep CTR/CVR predictors, and cold‑start paths, delivering up to 33 % spend growth, 6 % eCPM lift and lower latency while maintaining diversity.
The article examines how large language models can streamline advertising copy creation by addressing format diversity, creativity, and new media demands, detailing model evaluation, fine‑tuning of Chinese‑adapted LLMs—ultimately selecting QWen 1.5‑7B—and showing that deployment boosts copy quality, click‑through and conversion rates while outlining future personalization and data‑efficient scaling.
Ximalaya’s deep advertising conversion optimization advances from shallow to deep billing models by integrating OCPC dual‑bidding, full‑channel data assistance, and real‑time crowd premium to overcome data sparsity, long conversion delays, and cold‑start challenges, boosting advertisers’ ROI while managing platform risk and guiding future ROI‑protected bidding.
The Adaptive Domain Interest Network (ADIN) introduces a shared backbone with scenario‑specific subnetworks, domain‑specific batch normalization and SE‑Block attention to capture both commonalities and divergences across recommendation scenarios, and, combined with self‑supervised training, consistently outperforms baselines, delivering a 1.8% revenue lift in Alibaba’s display‑ad platform and now runs in production.
The Co‑Action Network replaces costly Cartesian‑product feature crossing with lightweight micro‑net‑based interaction units that share parameters across feature pairs, delivering comparable CTR prediction accuracy while cutting parameters to one‑tenth and boosting online latency, as proven in large‑scale advertising deployments.
The paper introduces a Meta‑Network based Multi‑Scenario Multi‑Task (M2M) model for Alibaba’s advertising merchants, combining a transformer‑driven backbone with scene‑aware meta‑learning modules to jointly predict spend, clicks and activity across diverse ad scenarios, achieving up to 27 % error reduction offline and over 2 % lifts in merchant activity and ARPU online.
The paper presents a guarantee‑based, distributed allocation framework for Alibaba’s off‑site brand contract ads that extends the SHALE algorithm with effect‑driven objectives and explicit over‑allocation constraints, solves dual variables via coordinate descent, and employs adaptive probability‑based pacing to meet volume guarantees while significantly boosting average CTR.
The paper presents a holistic algorithm‑system‑hardware co‑design for advertising deep‑learning inference, combining model pruning, approximate computing, kernel fusion, scheduling and PCIe transfer optimizations with GPU and NPU upgrades, achieving up to five‑fold speed‑up and significantly higher latency‑bounded QPS for large‑scale ad services.
At WSDM 2022, Alibaba’s advertising team presented four papers introducing a meta‑learning multi‑task multi‑scenario model for advertiser forecasting, a low‑cost Feature Co‑Action Network that boosts CTR prediction, an Adaptive Unified Allocation Framework that improves guaranteed display fulfillment and CTR, and a cooperative‑competitive multi‑agent auto‑bidding system that enhances both advertiser welfare and platform profit.
The paper proposes a unified solution for real‑time bidding in online display ads that formulates advertiser budget and KPI limits as a constrained linear program, derives a closed‑form optimal bidding function with m+1 parameters, and uses model‑free reinforcement learning to dynamically adjust those parameters, achieving superior traffic‑value capture in large‑scale deployment on Alibaba’s Taobao platform.
Alibaba's advertising platform combines knowledge distillation to score traffic, exposure forecasting via GBDT, and PID-based pacing to guarantee contracted impression volumes while improving CTR/CVR, handling delayed exposure and traffic selection, achieving near‑perfect delivery in large promotions.
Alibaba Mama’s advertising platform modernized its Tree‑based Deep Model by introducing a dual‑tower full‑library DNN with aggressive pre‑filtering and custom GPU TopK kernels, and a flattened‑tree model that retains beam search with multi‑head attention, while applying memory‑aware tricks such as attention swapping, softmax approximation, tiled‑matmul splitting, TensorCore batching, INT8 quantization and cache‑resident ad vectors, enabling multi‑fold latency reductions with minimal recall loss.