How Graph Embedding Boosts E‑Commerce Recommendations: GES & EGES Explained

Background

Alibaba’s personalized recommendation system faces billions of users, items, and interactions, forming a massive heterogeneous graph. Modeling this graph in a unified vector space can greatly simplify and enhance recommendation capabilities, yet mature graph‑embedding solutions for such scale were lacking.

This article introduces an innovative framework based on graph embedding to improve recommendation diversity and address cold‑start problems. It constructs a user‑behavior graph, applies random walks for virtual sampling to capture multi‑order interests, and incorporates side‑information‑based models, resulting in two vector aggregation algorithms: Graph Embedding with Side Information (GES) and Enhanced Graph Embedding with Side Information (EGES).

Base Graph Embedding Framework

The core architecture of graph embedding in Alibaba’s recommendation pipeline is illustrated below:

A directed weighted graph is built from user‑item interaction sequences, using transition probabilities derived from co‑occurrence frequencies to mitigate hotspot nodes. Random walks on this graph generate billions of multi‑order virtual samples for downstream deep learning. Sampled Softmax is employed for ultra‑large‑scale classification, optimizing the objective of maximizing node co‑occurrence.

GES and EGES Algorithms

Traditional collaborative filtering struggles with cold‑start items. GES extends the Skip‑Gram phase of graph embedding by jointly learning representations for nodes and their side information, merging multiple latent vectors into a final item embedding.

EGES further refines this by assigning different weights to various side‑information dimensions (e.g., brand, store) through a weighted pooling layer, improving the accuracy of vector fusion. The fusion formulas are shown below:

The EGES embedding network architecture is depicted as follows:

Incorporating side information brings similar items (same brand, store, etc.) closer in ranking, and enables effective embedding of fresh items without interaction history, thus solving the cold‑start problem.

Experimental Results

Extensive experiments on Alibaba’s internal dataset and the public Amazon dataset demonstrate significant performance gains. The algorithms were also deployed in Alibaba’s front‑page personalized recommendation, yielding observable improvements.

Visualization of shoe category embeddings shows clear clustering of items within the same sub‑category, confirming the quality of the learned vectors.

Cold‑start item recall examples illustrate that EGES captures generalized similarity via side information, effectively retrieving relevant items for newly introduced products.

System Deployment

EGES was launched before the 2017 Double‑11 shopping festival. The complete engineering architecture is shown below:

This article is a review of the paper “Billion‑scale Commodity Embedding for E‑commerce Recommendation in Alibaba,” presented at SIGKDD 2018.