How Alibaba’s Graph Embedding Boosts E‑Commerce Recommendations by 60%

Introduction

In the e‑commerce era, consumers are accustomed to recommendation systems, and Alibaba’s merchant division has launched the “thousands‑of‑people‑one‑face” intelligent shop module to let merchants pre‑select a candidate product set and let the algorithm personalize recommendations for each visitor.

The module addresses the difficulty of traditional matching in a marketplace with millions of active shops, where a single user’s behavior inside a specific shop is extremely sparse.

Scalable Graph Embedding Approach

Alibaba proposes a highly scalable graph‑embedding technique that learns low‑dimensional vectors for products. The vectors enable efficient computation of similarity between any two items with minimal storage, even when a user has never visited the shop before.

Method Details

The method assigns each node two embeddings – a source vector and a target vector – and uses the dot product Source(A)·Target(B) to model asymmetric similarity sim(A,B). Positive samples are drawn via Monte‑Carlo rooted PageRank walks, while negative samples are drawn randomly.

The training objective follows the Skip‑Gram with Negative Sampling formulation, optimizing a loss that approximates the logarithm of the rooted PageRank similarity.

Experimental Results

On a small internal dataset, the embedding method (named APP) outperforms traditional item‑to‑item metrics on link‑prediction AUC and top‑k recommendation tasks, especially when the test set contains many asymmetric edges.

When deployed in the “thousands‑of‑people‑one‑face” module, click‑through rate increased by 30 % and transaction volume by 60 % compared with the previous system.

References

Fogaras et al., “Towards scaling fully personalized PageRank”, Internet Mathematics, 2005.

Grover & Leskovec, “node2vec: Scalable feature learning for networks”, KDD, 2016.

Haveliwala, “Topic‑sensitive PageRank”, WWW, 2002.

Levy & Goldberg, “Neural word embedding as implicit matrix factorization”, NIPS, 2014.

Mikolov et al., “Distributed representations of words and phrases and their compositionality”, NIPS, 2013.

Perozzi et al., “DeepWalk: Online learning of social representations”, KDD, 2014.

Tang et al., “LINE: Large‑scale information network embedding”, WWW, 2015.

Liu et al., “PowerWalk: Scalable personalized PageRank via random walks”, CIKM, 2016.