How Alibaba Leverages Graph Embedding & Deep Learning for Double 11 Home‑Page Recommendations

Introduction

Alibaba's Double 11 mobile Taobao homepage hosts the largest traffic entry for the app, making its personalized recommendation scene a critical component of the overall traffic flow, distribution, and user interest discovery.

1. Business Technical Overview

The homepage personalization relies on Graph Embedding recall models and DeepCross&ResNet real‑time ranking models, built on top of search engineering systems such as Porsche&Blink, Rank Service, and Basic Engine.

2. Homepage Personalization Framework (MATCH recall + RANK sorting)

2.1 All Items Are Vectors – Graph Embedding Deep Recall Framework

To address long‑tail coverage and cold‑start problems, a Graph Embedding framework constructs a whole‑network behavior graph from user click sequences and applies deep random walks to generate multi‑hop virtual samples, expanding potential interest recall and improving diversity.

Graph Embedding projects complex heterogeneous networks into a low‑dimensional space, enabling unified vector representations for users, items, categories, and side information.

2.1.1 Graph Embedding – DeepWalk Algorithm

DeepWalk treats random walks on the graph as sentences and trains a Skip‑Gram model to learn node embeddings, maximizing co‑occurrence probabilities.

2.2 S³ Graph Embedding Model

The S³ model adds three innovations: Sequence Behavior, Sub‑graph partitioning, and Side‑information integration. It evolves from a naive version using swing similarity, to a Sequence+Side‑information version that incorporates user sessions, weighted directed graphs, and dynamic negative sampling, finally to a sub‑graph parallel training version that scales to billions of samples.

2.3 Deployment and Optimization in Recommendation Scenarios

Training runs on the XTensorflow platform with billions of sampled sequences, using GPU clusters for parallel embedding learning. The resulting embeddings are indexed for nearest‑inner‑product search on a GPU‑based online engine, replacing previous swing‑based i2i recall and delivering richer, more diverse results.

3. Deep Ranking Models Based on XTF

3.1 XTensorflow Overview

XTensorflow (XTF) is a distributed TensorFlow training and online scoring platform built on Porsche Blink, supporting rapid iteration and real‑time computation for high‑traffic homepage scenarios.

3.2 DeepResNet in AIOplus Scene

DeepResNet extends a Wide&Deep model with residual layers to mitigate gradient vanishing while learning high‑order features from sparse embeddings, improving CTR for the AIOplus card‑scene during Double 11.

3.3 DCN (Deep & Cross Network) for Main‑Hall Entrance

DCN combines a cross network for explicit high‑order feature crossing with deep layers for representation learning, achieving high precision in the limited‑slot main‑hall entrance scenario.

3.4 Union‑DeepModel

Building on the Wide&Deep paradigm, the Union‑DeepModel integrates various advanced architectures (WDL, PNN, DeepFM, NeuralFM, DCN, DeepResNet) into a unified TensorFlow Estimator‑based framework, enabling flexible deployment across diverse business scenarios.

Conclusion

The Alibaba recommendation team now serves billions of users daily across Taobao, Tmall, and other platforms, continuously improving traffic efficiency, user experience, and merchant engagement through AI‑driven personalization.