AI Powering NetEase Yanxuan: Supply Chain Forecasting, Personalization & Chatbots

Introduction

NetEase Yanxuan is an internet‑brand e‑commerce platform tightly integrated with manufacturing, covering the entire retail loop from product research and production to supply chain, sales, and after‑sale service. By deeply participating in the supply chain and leveraging NetEase's internet advantages, Yanxuan serves both manufacturers and consumers, enabling manufacturers to better understand customers and offering consumers higher‑value products.

Supply Chain Sales Forecasting

Accurate sales forecasting supports various planning tasks—demand, replenishment, allocation, and promotional planning—aiming to reduce inventory, prevent stock‑outs, and maintain a good user experience.

Key challenges include long forecasting horizons (2‑3 months or more), multiple sales channels (online marketplaces, offline stores, corporate customers), large promotional volatility, discontinuous data for new or iterated products, and multi‑granular demand (daily/weekly/monthly, SKU/SPU/category, region/warehouse/channel).

To address these, Yanxuan employs machine‑learning models that transform coarse‑grained activity plans into fine‑grained historical plans for training, use short‑term promotional plans at SKU level, and label new items with similar target users for model input.

Various models have been developed: point estimates, interval forecasts, event‑impact factors, and seasonal trend predictions, using statistical methods, XGBoost tree models, and exploratory TensorFlow deep models. Probabilistic sales‑distribution forecasts and seasonal predictions have been deployed online, with ongoing work on TensorFlow Probability to improve the probabilistic network.

Recent results show forecasting accuracy exceeding 65%, supporting thousands of SKUs across the supply‑chain and sales pipelines, handling tens of thousands of promotional events, and earning multiple technical patents.

Sales‑Side Personalization Algorithms

Personalization directly impacts user experience, transaction volume, click‑through rate (CTR), conversion rate (CVR), UV value, and overall GMV. Real‑time DeepCTR models capture user actions (e.g., searching “beef” and clicking a steak) and adjust ranking for subsequent similar queries.

TensorFlow facilitates feature engineering, multi‑task learning, and real‑time user preference modeling. A Wide&Deep architecture (Embedding & MLP) is used: embeddings for item IDs, category IDs, attribute IDs; attention‑based user vectors derived from behavior sequences; wide part creates high‑coverage cross features; deep part combines ID embeddings through deep networks.

Multi‑task learning adds CTR and CVR heads sharing the embedding layer, enabling a combined CTCVR loss for better exposure‑stage CVR estimation.

Online deep learning approximates real‑time preference detection by extracting user session interaction sequences, applying sequence embeddings, and maintaining multiple category‑level interest vectors per user for context‑aware re‑ranking.

Customer Service Chatbot

In e‑commerce, customers raise diverse, colloquial queries before and after purchase. The chatbot pipeline first identifies user intent, routes to appropriate sub‑modules, extracts fine‑grained semantics, matches against product and knowledge bases, and generates answers.

TensorFlow powers a multi‑level intent‑recognition model that uses current input, context, and historical behavior. Named‑entity recognition models extract product names, attribute names, and values, feeding a knowledge‑graph QA system.

For FAQ retrieval, a text‑matching and similarity model re‑ranks candidate answers, while a generative transformer‑based dialogue model produces human‑like responses.

Leveraging open‑source BERT and TensorFlow implementations, Yanxuan builds high‑accuracy text models with limited labeled data, meets online QPS requirements via GPU acceleration, and rapidly deploys via TensorFlow Serving.

Since 2016 Yanxuan transitioned from traditional machine learning to deep learning, fully embracing TensorFlow across supply, sales, and after‑sale services, with further expansion into upstream product development and manufacturing.