JD Digits Secures 16 Paper Acceptances at AAAI 2021, Showcasing Advances in Federated Learning, Spatio‑Temporal AI and Recommendation Systems

Recently, the top AI conference AAAI 2021 released its paper acceptance list, and JD Digits (京东数科) stood out with a record 16 accepted papers, spanning 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% (1,692 out of 9,034 submissions), JD Digits' high acceptance volume demonstrates its international‑level AI capabilities and its transition from laboratory research to practical deployments in smart cities, agriculture, commodities, retail, and AI robotics.

One highlighted work is the federated learning platform Fedlearn, which integrates cryptography, machine learning and blockchain to enable secure, intelligent, high‑efficiency data collaboration without moving raw data across institutions.

Two federated‑learning papers were accepted: "Secure Bilevel Asynchronous Vertical Federated Learning with Backward Updating" , proposing a novel vertical federated learning framework (VFB2) with three new algorithms (VFB2‑SGD, VFB2‑SVRG, VFB2‑SAGA) that improve efficiency and privacy; and "On the Convergence of Communication‑Efficient Local SGD for Federated Learning" , introducing a double‑compression error‑compensation mechanism that drastically reduces communication overhead for large‑scale models.

In the smart‑city domain, the paper "Traffic Flow Forecasting with Spatial‑Temporal Graph Diffusion Network" presents a heterogeneous graph neural network that models both temporal and spatial traffic patterns, outperforming methods that consider only local spatial relations.

Another contribution, "Robust Spatio‑Temporal Purchase Prediction via Deep Meta Learning" , offers the STMP model—a meta‑learning based spatio‑temporal multi‑task deep generative framework—for accurate sales forecasting during shopping festivals.

For recommendation, JD Digits released "Graph‑Enhanced Multi‑Task Learning of Multi‑Level Transition Dynamics for Session‑based Recommendation" , which uses heterogeneous attention and cross‑session graph learning to capture both short‑term and long‑term item transitions; and "Knowledge‑aware Coupled Graph Neural Network for Social Recommendation" , which incorporates knowledge‑based item relations to improve robustness against user sparsity.

In risk management, the paper "The Causal Learning of Retail Delinquency" applies double machine learning to estimate the causal effect of credit limits on user risk, addressing survivor bias and aiding more scientific credit‑granting strategies.

The 16 accepted papers underscore JD Digits' rapid research output—nearly 150 papers across AAAI, IJCAI, CVPR, KDD, NeurIPS, ICML by January 2021—and its heavy investment in R&D (≈70% of staff, ~16% of revenue in H1 2020), positioning it alongside global internet giants.

Looking forward, JD Digits will continue to leverage AI to drive industry digitalization, integrating technology, industry, and ecosystem to provide comprehensive digital services.