Federated Learning in E‑commerce Marketing: JD.com’s 9N‑FL Platform Overview and Practices

Data is the oil of AI, but high‑quality, multi‑dimensional data is fragmented into isolated silos due to privacy, commercial secrecy, and regulatory constraints, limiting the training of more powerful models.

Federated learning (FL) addresses this challenge by enabling multiple participants to collaboratively train models without exposing raw data, thus unlocking new business scenarios while preserving privacy.

The JD Retail‑Technology and Data Center team built the 9N‑FL platform, a large‑scale industrial FL system, and applied it to e‑commerce marketing. The article is organized into six parts: FL background, FL introduction, application scenarios, 9N‑FL details, privacy protection, and planning summary.

Key FL characteristics include multiple participants with complementary data, a training mode where data never leaves the local domain, cross‑domain deployment of compute resources, secure encrypted transmission of model updates, and model performance that closely matches centrally trained models.

FL is classified into three main types: horizontal FL (same feature space, different samples), vertical FL (same samples, different features), and federated transfer learning. JD’s 9N‑FL supports all three, with a simple illustration of horizontal and vertical architectures.

In the e‑commerce advertising scenario, JD’s media side provides user interest tags, while JD supplies commercial interest and conversion labels. By aligning session IDs, both sides jointly train models, achieving a 15%+ increase in revenue.

Privacy protection techniques employed include privacy‑preserving set intersection (PSI) for sample alignment, homomorphic encryption, secret sharing, garbled circuits, and differential privacy. The article details a pipeline‑based PSI solution that scales to billions of samples without performance loss.

9N‑FL’s innovations comprise end‑to‑end FL workflow (sample matching to model training/prediction), support for massive scale (hundreds of billions of samples, terabytes of data), distributed asynchronous framework with failover and congestion control, and high‑performance network scheduling.

The platform, built on JD’s 9N machine‑learning stack (TensorFlow core, Kubernetes orchestration), supports classic algorithms (LR, tree) and neural networks, offering encrypted tree models for high‑privacy domains and unrestricted neural network training for large‑scale scenarios.

Overall, federated learning breaks data silos, enabling secure multi‑party collaboration and becoming a cornerstone for future AI development.