Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba
EGES integrates side information and reaches 0.9746 Taobao AUC, a 10.8% gain over BGE.
Key Findings
Methodology
The paper builds a weighted directed item graph from consecutive actions within one-hour user sessions, then learns representations with DeepWalk, Skip-Gram, and negative sampling. BGE uses behavior alone; GES averages item and side-information embeddings; EGES learns positive, normalized exponential weights for different attributes and optimizes a binary logistic objective for context prediction.
Key Results
- On Amazon, BGE, GES, and EGES obtain AUCs of 0.9327, 0.9575, and 0.9700, respectively. EGES improves 4.00% over BGE, showing that side information and learned aggregation substantially improve link prediction.
- On Taobao, BGE reaches 0.8797, GES 0.9704, and EGES 0.9746, giving EGES a 10.8% relative improvement over BGE. It also exceeds LINE(1st) at 0.9100 and LINE(2nd) at 0.9411.
- A seven-day Mobile Taobao A/B test in November 2017 shows that GES and EGES consistently outperform the production item-CF baseline and BGE in homepage CTR; EGES remains above GES throughout the test.
Significance
The work demonstrates that graph representation learning can serve a production recommender operating at roughly one billion users and two billion items. It addresses a practical combination of scale, extreme sparsity, sequential behavior, and continuously arriving products. By separating fast matching from deep ranking, it provides a scalable route to improve candidate quality without requiring an expensive full user-item model at retrieval time.
Technical Contribution
The paper contributes a session-based directed graph construction, Base Graph Embedding (BGE), Graph Embedding with Side Information (GES), and Enhanced Graph Embedding with Side Information (EGES). EGES forms H_v=Σe^{a_v^s}W_v^s/Σe^{a_v^s}, then trains item, attribute, and weight parameters with negative-sampling logistic loss, combining higher-order behavior structure with cold-start generalization.
Novelty
Relative to item-based CF, which mainly exploits co-occurrence, and LINE, which models graph proximity, the paper inserts heterogeneous side information directly into graph-embedding training. Its fundamental innovation is learned, type-specific aggregation rather than merely appending features, producing an industrially deployable representation for sparse and cold-start catalogs.
Limitations
- The evaluation emphasizes link-prediction AUC and CTR, without reporting long-term GMV, conversion, retention, fairness, or statistical significance; higher CTR does not necessarily imply higher customer value.
- Random walks, negative sampling, and full-catalog updates remain costly at billion scale; performance also depends on clean attributes, the one-hour window, and stable learned weights.
Future Work
Promising directions include temporal graphs, joint user-item representations, incremental updates, and multimodal attributes. Future systems should evaluate exposure bias, popularity amplification, privacy, transfer across scenarios, and objectives beyond clicks, including purchase probability, long-term value, diversity, and fairness.
AI Executive Summary
Taobao contains approximately one billion users and two billion items, making conventional collaborative filtering difficult to scale and unreliable for sparse or newly listed products. The paper focuses on the matching stage of a two-stage recommender: retrieve similar candidates first, then rank them with a deep neural network. Item co-occurrence alone ignores sequence and cannot represent items with no interaction history.
The authors build a weighted directed graph from consecutive actions within one-hour sessions. BGE applies DeepWalk random walks followed by Skip-Gram and negative sampling. GES averages item, category, brand, and other side-information embeddings, while EGES learns different exponential weights for different information types. Thus, representations combine behavioral proximity with attribute-based generalization.
With 160-dimensional embeddings, walk length 10, 20 walks per node, and context window 5, EGES reaches AUC 0.9700 on Amazon and 0.9746 on Taobao, improving over BGE by 4.00% and 10.8%. It surpasses LINE variants as well. In a seven-day November 2017 online test, EGES consistently improves homepage CTR over the production item-CF baseline, BGE, and GES. Deployed on Alibaba’s XTensorflow platform, the system demonstrates a credible path from graph learning to industrial retrieval, although long-term business effects, bias, and real-time adaptation remain open challenges.
Deep Analysis
Background
Taobao had roughly one billion users and two billion items in 2017; the Mobile Taobao homepage contributed about 40% of recommendation traffic. Collaborative filtering, content methods, and deep models work at smaller scales but struggle with sparse interactions, rapid catalog growth, and memory or latency limits. The paper builds on word2vec, DeepWalk, and LINE to bring graph representation learning into industrial item matching.
Core Problem
The task is to compute item-item similarities for candidate generation, not final ranking. The system must handle billions of nodes, drifting user interests, sparse co-occurrence, and newly listed items without interactions. Full histories are expensive and mix obsolete interests; ordinary CF misses direction, sequence, and higher-order relations.
Innovation
First, consecutive events in one-hour sessions form a weighted directed graph that preserves transitions. Second, BGE uses DeepWalk to capture multi-hop similarity. Third, GES averages item and side-information embeddings for sparse or unseen items. Fourth, EGES learns attribute-specific weights instead of assuming category, brand, shop, and other signals contribute equally. The methods are implemented on XTensorflow for production-scale processing.
Methodology
- �� Cleaning: remove clicks with less than one second of dwell time, suspected spam users, and item identifiers whose semantics changed through repeated updates.
- �� Graph construction: connect consecutive session items; edge weight equals global transition frequency. Random-walk transition is P(v_j|v_i)=M_ij/Σ_jM_ij.
- �� BGE: generate walks of length 10, 20 walks per node, then train 160-dimensional Skip-Gram embeddings with window 5 and negative sampling.
- �� GES: average item and n side-information vectors, H_v=(1/(n+1))ΣW_v^s.
- �� EGES: compute H_v=Σe^{a_v^s}W_v^s/Σe^{a_v^s}; update embeddings and weights using logistic loss and stochastic gradients.
Experiments
For link prediction, one third of edges are randomly removed as positives and an equal number of nonexistent pairs are sampled as negatives; AUC is the metric. Amazon Electronics has 300,150 nodes, 3,740,196 edges, and three information types. Taobao has 2,632,379 nodes, 44,997,887 edges, and twelve types; sparsity is 99.58% and 99.943%, respectively. Baselines are BGE, LINE(1st), LINE(2nd), GES, and EGES. Online testing compares them with production item-CF.
Results
Amazon AUCs are 0.9327 for BGE, 0.9554 for LINE(1st), 0.9575 for GES, and 0.9700 for EGES. Taobao scores are 0.8797, 0.9100, 0.9704, and 0.9746, respectively, with LINE(2nd) at 0.9411. GES and EGES beat all BGE and LINE comparisons. The larger Taobao gain is attributed to twelve useful side-information types. Online CTR ranks EGES above GES, BGE, and Base throughout seven days.
Applications
The method directly supports homepage recommendations, similar-item retrieval, advertising candidate generation, and new-product exposure. Required infrastructure includes clean session logs, reliable catalog attributes, offline graph training, and vector-neighbor serving. Because retrieval uses item vectors before neural ranking, platforms can improve candidate quality while keeping online latency manageable.
Limitations & Outlook
The approach assumes a fixed one-hour window, trustworthy attributes, and relatively stable item semantics. It does not fully model seasonality, rapid interest shifts, images, text, or evolving graph structure. Random walks, negative sampling, and full-vector refreshes can be expensive at two-billion-item scale. The paper omits training cost, latency, confidence intervals, and detailed CTR values, and does not study popularity bias, homogenization, or long-term value.
Plain Language Accessible to non-experts
Imagine a gigantic department store whose shelves change every hour. The store watches shoppers’ routes: if many people examine a phone and then headphones, it draws a directed path between those products, making the path thicker when the pattern repeats.
The first system sends a small explorer along these paths. By wandering repeatedly, it discovers products that may not be directly connected but tend to appear in the same shopping journeys. The second system also reads labels such as brand, category, and shop. A new product with no visitors can still be placed near older products sharing those labels. The third system learns which labels matter most rather than treating every label equally.
The resulting neighbors are handed to a separate clerk that chooses what each shopper should actually see. On Taobao, this method achieved 0.9746 AUC, versus 0.8797 for the basic graph model, and it improved online clicks for seven consecutive days. It is therefore like a self-learning store map. Yet wrong labels, changing tastes, and excessive focus on popular products can still mislead the map.
ELI14 Explained like you're 14
Picture a game shop with two billion items. Players only inspect a few things, so comparing every player with every item would be insanely slow! The paper watches what players do in sequence. If lots of people check item A and then item B, the system draws a road from A to B.
A tiny digital walker travels these roads again and again. After enough trips, it learns that two items can be related even when they are not next-door neighbors. This is the DeepWalk idea: each item gets a position made of numbers, and nearby positions mean similar shopping behavior.
But what about a brand-new sword nobody has clicked yet? The system checks its labels—brand, category, store, price, and so on. GES mixes those clues equally. EGES is smarter: it learns that brand may matter a lot for phones, while style may matter more for clothes. It uses the clues differently instead of treating them like identical teammates.
The result is impressive: EGES scores 0.9746 on the Taobao test, while BGE scores 0.8797. It also gets higher homepage click-through rates for seven days. In short, the system first learns from everyone’s shopping routes, then uses labels to help new items join the map. Cool, right? Still, it needs clean data and must avoid showing only famous products!
Glossary
Graph Embedding
A way to turn graph nodes into low-dimensional vectors so structurally related nodes are close together. Technically, it learns a mapping from nodes to R^d while preserving neighborhood information.
The paper represents Taobao items as vectors and uses dot products for similarity-based retrieval.
DeepWalk
A graph representation method that generates random-walk sequences and trains Skip-Gram on them. It captures higher-order proximity beyond immediate edges.
DeepWalk is the sequence-generation foundation of BGE, GES, and EGES.
Skip-Gram
A representation-learning objective that predicts context elements from a center element. Negative sampling makes this probability objective computationally practical.
It learns item and context embeddings from random-walk sequences.
Side Information
Attributes external to interaction history, such as brand, category, shop, price, age, or gender. These features provide signals for sparse and unseen items.
GES averages them and EGES learns their relative contributions.
Cold Start
The difficulty of recommending a new user or item with little or no historical interaction. E-commerce suffers especially because products arrive continuously.
Side-information aggregation gives cold-start items usable initial embeddings.
Open Questions Unanswered questions from this research
- 1 The paper does not disclose complete training cost, serving latency, or refresh strategy for two billion items. Reproducible system benchmarks are needed to establish how EGES behaves under real-time catalog updates.
- 2 It remains unclear whether CTR gains produce higher purchases, GMV, retention, or user welfare. Causal evaluation should examine exposure bias, popularity amplification, diversity, and long-term outcomes.
Applications
Immediate Applications
Homepage similar-item retrieval
An e-commerce platform can construct session graphs, train EGES vectors, retrieve nearest neighbors for items in a user’s history, and pass candidates to its existing ranker. Clean logs, catalog attributes, offline training, and a vector-search service are the main prerequisites.
New-product cold-start exposure
A newly listed product can receive an initial embedding from brand, category, shop, price, and related attributes before it accumulates clicks. It can then enter similar-item, advertising, and campaign candidate pools immediately.
Long-term Vision
Real-time multi-objective recommendation
A future platform could combine temporal decay, text and image signals, user histories, and long-term feedback while optimizing clicks, purchases, retention, diversity, and fairness. The main obstacles are computation, bias control, and reliable online evaluation.
Abstract
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.