Alibaba Retail's Intelligent Recommendation System: Business Background, Architecture, Matching and Ranking Models

Business Background – Retail Pass is Alibaba’s B‑class business unit that provides an end‑to‑end solution for offline community stores, enabling them to source quality goods, upgrade convenience services, and connect brand merchants with millions of small shops through a B2B2C distribution model.

The APP offers three core recommendation scenes: a large‑scale promotion hall (e.g., 618 event), daily channels (e.g., trending foods, selected items, limited‑time offers), and a "Guess You Like" module that appears on the home page, order page, bundle page, product detail page, and payment‑success page, covering the entire purchase journey.

Technical Architecture – The recommendation flow starts with real‑time data stored in ABFS. TPP (The Personalization Platform) queries ABFS and UPS to infer user intent, the BE (Basic Engine) performs recall, and RTP (Real‑time Prediction) scores and re‑ranks items before display. Offline and online model training, as well as parameter tuning, are handled by PAI and Porsche, supporting both T+1 and real‑time updates.

Matching Models

Item‑CF – Uses eTREC and Swing algorithms to compute item similarity on a user‑object bipartite graph, providing robust noise resistance and mitigating long‑tail effects.

Graph Embedding I2I – Applies weighted random walks on the user‑item graph and trains embeddings with Skip‑Gram (DeepWalk, Node2Vec). Top‑K similar items are retrieved for order‑page recommendations.

Side‑Information Enhanced Embedding – Incorporates product side information to improve generalization and address cold‑start problems.

User‑CF with LBS – Calculates shop similarity using LBS attributes (business circle, city tier, street, store size) and purchase‑sequence vectors (Doc2Vec + SimHash), then combines both scores for a hybrid recommendation.

Ranking Models – Consist of coarse‑ranking, fine‑ranking, and re‑ranking stages. Fine‑ranking models target business goals such as CTR, CVR, and SKU diversity. Feature categories include discrete IDs (shop, product, context), continuous statistical features, and sequential features (purchase cycles, query logs).

The model evolution progressed from linear models (LR, GBDT) to deep models (FM, Wide&Deep, DeepFM) and now explores attention‑based and Transformer architectures.

APP Shelf Recommendation – The shelf mimics the physical layout of a small store, using LBS‑driven group‑item recommendation and a combination of online consumer data (Alipay, Taobao, Ele.me) and offline data (POS, wholesale channels) to generate profit‑maximizing assortments.

Summary – The article outlines the end‑to‑end pipeline of Alibaba Retail’s intelligent recommendation system, covering business challenges, algorithmic architecture, matching and ranking innovations, and practical insights for B2B2C e‑commerce scenarios.

References

The Link Prediction Problem for Social Networks (2004)

DeepWalk: Online Learning of Social Representations (2014)

Node2vec: Scalable Feature Learning for Networks (2016)

Billion‑scale Commodity Embedding for E‑commerce Recommendation in Alibaba (2018)

Multi‑Interest Network with Dynamic Routing for Recommendation at Tmall (2019)

Behavior Sequence Transformer for E‑commerce Recommendation in Alibaba (2019)

Deep Crossing: Web‑Scale Modeling without Manually Crafted Combinatorial Features (2016)

Wide & Deep Learning for Recommender Systems (2016)

Factorization Machines (ICDM 2010)

DeepFM: a factorization‑machine based neural network for CTR prediction (2017)

Attention Is All You Need (2017)