Source-linked AI summary
Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba
Jizhe Wang, Pipei Huang, Huan Zhao, Zhibo Zhang, Binqiang Zhao, Dik Lun Lee
TL;DR
Taobao’s billion-scale recommendation data poses scalability, sparsity, and cold-start challenges. The paper constructs item graphs from user behavior, learns item embeddings, and incorporates side information to improve matching. Offline and online experiments report improved recommendation accuracy and CTRs, while deployment demonstrates feasibility in live production.
Problem
Taobao’s recommendation system must handle one billion users and two billion items while addressing scalability, sparsity, and cold start.
Method
The paper constructs an item graph from user behavior, learns item embeddings for similarity-based matching, and incorporates side information for sparse and cold-start items.
Results
Offline experiments find side-information methods more effective than methods without side information, while online A/B tests report improved CTRs in Taobao production.
Takeaways & Limitations
Graph embedding with side information provides an effective and feasible approach for billion-scale item matching in Taobao’s live recommendation environment.
Takeaways & Limitations
The base graph-embedding formulation relies on an independence assumption for context nodes in behavior sequences.
Abstract
from arXiv · showhide
Recommender systems (RSs) have been the most important technology for increasing the business in Taobao, the largest online consumer-to-consumer (C2C) platform in China. The billion-scale data in Taobao creates three major challenges to Taobao's RS: scalability, sparsity and cold start. In this paper, we present our technical solutions to address these three challenges. The methods are based on the graph embedding framework. We first construct an item graph from users' behavior history. Each item is then represented as a vector using graph embedding. The item embeddings are employed to compute pairwise similarities between all items, which are then used in the recommendation process. To alleviate the sparsity and cold start problems, side information is incorporated into the embedding framework. We propose two aggregation methods to integrate the embeddings of items and the corresponding side information. Experimental results from offline experiments show that methods incorporating side information are superior to those that do not. Further, we describe the platform upon which the embedding methods are deployed and the workflow to process the billion-scale data in Taobao. Using online A/B test, we show that the online Click-Through-Rate (CTRs) are improved comparing to the previous recommendation methods widely used in Taobao, further demonstrating the effectiveness and feasibility of our proposed methods in Taobao's live production environment.
1 INTRODUCTION
Taobao’s billion-scale recommendation system must address scalability, sparsity, and cold start while helping users find relevant items quickly. The paper proposes graph-embedding methods for matching items and reports improved production performance.
- Challenges: Taobao’s recommendation system serves one billion users and two billion items, creating scalability, sparsity, and cold-start challenges.Users interact with few items, while millions of new items are uploaded each hour without behavioral data.
- Recommendation framework: The matching stage computes pairwise item similarities from user behavior and generates candidates for later ranking.Taobao uses a two-stage framework: matching followed by deep-neural-network ranking.
- Graph embedding methods: BGE learns item embeddings from an item graph constructed from users’ behavior history, with similarities computed by embedding-vector dot products.Graph random walks capture higher-order item similarities beyond direct co-occurrence.
- Graph embedding methods: GES and related methods incorporate side information such as category, brand, and price to improve embeddings for sparse or cold-start items.A weighting mechanism accounts for different contributions from multiple side-information types.
- Contributions: The paper introduces three embedding methods—BGE, GES, and EGES—and evaluates GES and EGES against BGE and other embedding methods.The methods are designed to learn embeddings for two billion Taobao items.
- Results and deployment: The proposed framework significantly improves Mobile Taobao App recommendation performance while meeting training-efficiency and instant-response requirements during Double-Eleven Day.The system is deployed on Alibaba’s XTensorflow platform for billion-scale users and items.
2 FRAMEWORK
The framework constructs a weighted directed item graph from session-based user behavior, learns item embeddings through random walks and Skip-Gram, and incorporates side information to improve cold-start representations.
- Item-graph construction: Session-based behavior uses a one-hour window to capture sequential preferences while limiting computation and accommodating users’ changing interests.Only behaviors within the window are selected before constructing the graph.
- Item-graph construction: Directed edges connect consecutively interacted items, with weights equal to transition frequencies across all users’ behavior histories.The graph aggregates collaborative behavior into a weighted representation of item transitions.
- Base graph embedding: DeepWalk generates node sequences by random walk on the weighted graph, then Skip-Gram learns embeddings by maximizing node co-occurrence probabilities.The random walk uses outlink neighbors and transition probabilities derived from edge weights.
- Base graph embedding: The resulting embeddings capture higher-order similarities in behavior sequences that previous collaborative-filtering methods ignore.This embedding process provides low-dimensional representations for Taobao’s items.
- Side-information embedding: GES incorporates item and side-information embeddings, such as category, shop, and price, to improve representations for cold-start items.The method assumes items with similar side information should be closer in embedding space.
- Side-information embedding: EGES replaces equal aggregation with a weighted average because different side-information types contribute differently to item co-occurrence.The framework assigns distinct weights to the item itself and each side-information type.
3 EXPERIMENTS
The experiments evaluate graph-embedding methods offline and online, showing that side information and weighted aggregation improve recommendation quality. Case studies further examine embedding structure, cold-start recommendations, and learned side-information weights.
- Evaluation setup: The experiments evaluate link prediction offline and online CTR on the Mobile Taobao App, followed by real-world case studies.The offline task uses link prediction; the online goal is homepage Click-Through-Rate.
- Offline evaluation: GES and EGES outperform BGE and both LINE variants in AUC on Amazon and Taobao.The reported comparison attributes the improvement to incorporating side information.
- Offline evaluation: Taobao shows a larger performance gain from side information than Amazon, while Amazon shows a larger EGES-over-GES gain.The Taobao performance is reported as already very good at 0.97, limiting EGES’s visible improvement there.
- Online A/B test: EGES and GES consistently outperform BGE and Base in online CTR, while EGES consistently exceeds GES.Base is the previously deployed item-based CF method, and EGES uses weighted rather than average aggregation.
- Embedding visualization: EGES embeddings group shoes by category, place badminton and table-tennis shoes closer than football shoes, and support more relevant cross-category recommendations.The embeddings are projected into two dimensions using PCA, with colors representing categories.
- Cold-start items: For cold-start items, averaging side-information embeddings enables retrieval of similar existing items, with shops described as especially informative.The case study annotates the side-information types connecting retrieved items to each cold-start item.
- Side-information weights: Different items have different side-information weight distributions, supporting item-specific contributions to the final EGES representation.The weights are extracted from the learned weight matrix A for eight items across categories.
4 SYSTEM DEPLOYMENT AND OPERATION
Taobao’s recommendation platform separates online serving from offline graph-embedding computation. Offline logs produce item graphs and embeddings, while the online subsystem retrieves and ranks candidates for app users.
- Platform architecture: The recommending platform consists of online and offline subsystems.The online subsystem includes Taobao Personality Platform and Ranking Service Platform.
- Online workflow: When a user launches the Mobile Taobao App, TPP retrieves offline candidates and RSP ranks them with a fine-tuned deep neural network.The ranked results are returned to TPP.
- Data collection: User behaviors during Taobao visits are collected and saved as log data for offline processing.These logs supply the behavioral input to the offline subsystem.
- Offline workflow: Graph embedding methods are implemented and deployed in the offline subsystem of the recommending platform.The deployment section describes the platform first and then the modules relevant to graph embeddings.
- Offline workflow: The offline pipeline retrieves recent three-month logs, applies anti-spam processing, and constructs the item graph from the remaining behavior data.The remaining logs contain about 600 billion entries.
5 RELATED WORK
Related work covers graph embedding methods, graph embedding with side information, and graph embedding for recommender systems. It frames the paper’s approach against topology-only representations and recommendation models built on embeddings.
- Scope: The paper’s related-work scope includes graph embedding, side-information-enhanced graph embedding, and graph embedding for recommender systems.These are the three areas explicitly reviewed in the section.
- Graph embedding: Graph embedding methods are grouped into factorization, deep learning, and other broad methodological categories.The related-work discussion presents graph embedding as a general network representation method applied across real-world applications.
- Side information: Topology-only graph embeddings are described as vulnerable to sparsity and cold-start problems, motivating the use of side information.Prior work commonly assumes that nodes with similar side information should be closer in embedding space.
- Graph embedding for recommendation: Graph embeddings are used in recommender systems with downstream prediction models, including linear models and factorization machines.Prior recommendation work also learns user and item embeddings under supervision from heterogeneous-network structures.
6 CONCLUSION AND FUTURE WORK
The paper presents graph-embedding methods and deployment infrastructure for Taobao’s scalability, sparsity, and cold-start challenges. Side information improves offline recommendation accuracy, while online CTR results support the methods’ effectiveness and feasibility in production.
- Graph embedding methods address Taobao’s scalability, sparsity, and cold-start challenges.
- Incorporating side information into graph embedding improves recommending accuracy in offline experiments.
- Online CTRs demonstrate the effectiveness and feasibility of the proposed methods in Taobao’s live production.
- Real-world cases show graph embeddings clustering related items from user behavior and handling cold-start items with side information.
- The paper details training platforms and the recommendation workflow used to address scalability and deployment issues.