How Alibaba’s Deep Interest Network Powers Personalized Shopping for 400 Million Users

Alibaba Group Vice President Gu XueMei delivered a speech titled “Realizing Personalized Shopping for 400 Million People” at the 40th International ACM SIGIR conference.

Recent technological advances have made instant consumption in niche interest scenarios possible, such as women focusing on overseas goods or beauty products, and men on sports gear or electronics. Alibaba therefore needs to deeply understand user interests and intents, treating user and product information as equally critical.

As product catalogs grow and user demands evolve, e‑commerce platforms no longer present data in a simple tree structure; instead, they use knowledge graphs to link users, items, content, and their relationships.

Gu cited that before personalizing the 2015 Double‑11 event, the user churn rate in a non‑personalized venue was about 50 %. After opening a personalized venue, churn fell below 10 %, demonstrating that leveraging big data and large models to mine user interests is an inevitable trend for e‑commerce.

Alibaba, as one of the world’s largest e‑commerce platforms, feels a responsibility and advantage to lead lifestyle upgrades by delivering personalized pages for each user.

The three services contributing most to personalized shopping are search, recommendation, and advertising, all of which present content aligned with user interests to boost conversion.

Analysis of historical user behavior revealed two key metrics affecting click‑through prediction accuracy: diversity (users interested in many categories) and partial relevance (only some data can predict preferences, e.g., sunglasses recommended alongside swimwear but not books).

Inspired by attention models used in machine translation, Alibaba modified a basic multilayer fully‑connected neural architecture to propose the Deep Interest Network (DIN) structure.

Alibaba is also building its own industrial‑grade deep learning framework because the market lacks a ready‑made large‑scale framework tailored for personalized shopping, and the long, complex user journey in e‑commerce demands continual algorithmic and infrastructure improvements.

Currently, using DIN has yielded noticeable gains in ad click‑through rate prediction, and Alibaba plans to further refine the architecture to incorporate human knowledge, priors, instability handling, scalability, and basic reasoning.

In this rapidly evolving era, with new products and growing user expectations daily, Alibaba aims to continuously optimize its technologies to create a smarter market and a better world for users.