How JD Tech’s Breakthrough AI Papers Dominated AAAI 2021

Recently, the top international AI conference AAAI 2021 (the 35th AAAI) officially began, and JD Tech Group stood out with a record‑breaking 21 accepted papers covering computer vision, federated learning, adversarial learning, deep learning, sequential and social recommendation, graph neural networks, causal inference for risk management, and spatio‑temporal AI for smart cities.

With an overall acceptance rate of only 21%, JD Tech’s high paper count is notable, reflecting the company’s shift from pure laboratory research to practical deployments in smart cities, commodities, retail, agriculture, and AI robotics, accelerating industry digitalization.

AAAI (Association for the Advancement of Artificial Intelligence) is one of the oldest and most comprehensive top‑tier AI conferences, attracting massive submissions worldwide; it is classified as an A‑class conference in China’s ranking of international AI venues.

AAAI 2021 co‑chair Kevin Leyton‑Brown noted on Twitter that this year’s submissions reached an “astonishingly high technical level,” with 9,034 submissions, 7,911 reviewed, and only 1,692 papers accepted (21% acceptance).

JD Tech’s Self‑Developed Federated Learning Platform Breaks Data Silos

Addressing the challenge of “data islands,” JD Tech’s FedLearn platform integrates cryptography, machine learning, and blockchain to create a secure, intelligent, and efficient linking platform that enables multi‑party data collaboration without transmitting raw data.

Two JD Tech papers were accepted at AAAI 2021 in this area. The paper “Secure Bilevel Asynchronous Vertical Federated Learning with Backward Updating” proposes a novel vertical federated learning framework (VFB2) combining backward updating and asynchronous parallelism, along with three new algorithms (VFB2‑SGD, VFB2‑SVRG, VFB2‑SAGA) that enhance collaborative model training while preserving data privacy.

The paper “On the Convergence of Communication‑Efficient Local SGD for Federated Learning” introduces a new communication‑efficient distributed stochastic gradient algorithm that uses error‑compensated double compression to dramatically reduce communication overhead in large‑scale federated training.

Spatio‑Temporal AI Powers Smart City Operating Systems

JD Tech focuses on “spatio‑temporal AI” and presented several papers at AAAI 2021, such as “Traffic Flow Forecasting with Spatial‑Temporal Graph Diffusion Network,” which employs a heterogeneous graph neural network to model traffic flow across time and space, outperforming methods that consider only local spatial relationships.

The “Robust Spatio‑Temporal Purchase Prediction via Deep Meta Learning” paper proposes the STMP model, a meta‑learning based spatio‑temporal multi‑task deep generative model for retail sales forecasting during shopping festivals, enabling merchants to anticipate demand spikes.

JD Tech’s smart‑city operating system integrates a spatio‑temporal data engine, AI core, federated‑learning digital gateway, and visualization platform, achieving 10‑100× faster data processing than traditional platforms and reducing AI model development time from two years to two days.

Deployed in Xiong’an and Nanjing’s Nanjing‑based smart‑city command center, the system aggregates billions of data points from dozens of departments, providing real‑time dashboards for traffic, public safety, and environmental monitoring.

Advanced Recommendation Techniques for Precise Marketing

JD Tech’s papers on recommendation include “Graph‑Enhanced Multi‑Task Learning of Multi‑Level Transition Dynamics for Session‑based Recommendation,” which introduces a heterogeneous attention mechanism and cross‑session graph learning to capture both short‑term and long‑term item relationships.

The “Knowledge‑aware Coupled Graph Neural Network for Social Recommendation” paper incorporates product knowledge graphs into social recommendation, improving user preference modeling and mitigating sparsity issues.

Computer Vision, Video Learning, and Cross‑Modal Pre‑Training

JD Tech contributed papers such as “Exploiting Relationship for Complex‑scene Image Generation,” which uses relational semantics to guide multi‑object image synthesis, and “SeCo: Exploring Sequence Supervision for Unsupervised Representation Learning,” a video self‑supervised learning method that outperforms ImageNet‑pretrained models.

Another paper, “Scheduled Sampling in Vision‑Language Pretraining with Decoupled Encoder‑Decoder Network,” proposes a novel decoupled transformer architecture for both vision‑language understanding and generation tasks.

Causal Learning for Risk Management

The paper “The Causal Learning of Retail Delinquency” addresses causal effects of credit limits on user risk, employing double machine learning to correct survivor bias and produce unbiased estimators for strategy‑risk relationships, aiding more scientific credit decisions.

By January 2021, JD Tech had published nearly 350 papers across top AI conferences (AAAI, IJCAI, CVPR, KDD, NeurIPS, ICML, ACL, ICASSP) and won 19 world‑first awards, while collaborating with institutions like Stanford and USTC to foster AI talent and industry‑academic integration.