How IntentGC Scales Graph Convolution for Billion‑Node Recommendation Systems

1. Introduction

Recommendation systems typically model user‑item preferences as edges in a graph. In large‑scale e‑commerce platforms such as Taobao, the graph can contain billions of nodes and hundreds of billions of edges, making computation and inference extremely challenging. Network embedding methods learn low‑dimensional representations of nodes, enabling recommendation in a low‑dimensional space. This paper, published at KDD 2019, proposes IntentGC, a scalable graph convolution framework that fuses explicit user‑item interactions with rich heterogeneous information (search queries, visited shops, brand preferences, attribute preferences) to improve recommendation performance.

2. Problem Definition

In the e‑commerce scenario, the task is to recommend items to users. Historical user behavior forms a heterogeneous information network (HIN), an undirected graph where

denotes nodes,

denotes edges, and

represents all node types. User nodes and item nodes may be labeled (an existing edge) or unlabeled (no edge). The recommendation problem is cast as a link‑prediction task on the HIN: given the HIN constructed from historical behavior, predict future user‑item edges.

3. Model Design

The proposed model integrates multiple heterogeneous signals into a large‑scale graph convolution learning algorithm. It uses a bipartite heterogeneous graph, and the loss is designed with a triplet objective to balance explicit user preferences and auxiliary information. The learning process is semi‑supervised, leveraging abundant unlabeled data in the e‑commerce ecosystem.

Network Translation : Heterogeneous nodes are translated into homogeneous user‑user or item‑item relations based on second‑order similarity; if two users share many identical auxiliary connections, they are considered similar, enabling the encoding of heterogeneous semantics into homogeneous edges.

Fast Convolution Network (IntentNet) : Traditional GCN suffers from exponential complexity on large graphs. IntentNet reduces this by (1) sparsifying the convolution to activate only the most relevant neurons (channel‑shared vector learning) and (2) decoupling high‑order propagation into separate graph‑view and node‑view training modules. The vectorized convolution function is illustrated below: The formulas (3) and (4) encode node‑self and neighbor weights, which are shared across the graph.

Dual Convolution : To obtain precise user and item representations, separate convolutions are applied to user nodes and to item (plus negative sampling) nodes. Their outputs are projected into a common semantic space via a dense layer, and a triplet loss aligns user‑item pairs while pushing apart negative samples, yielding more accurate heterogeneous embeddings.

4. Experiments

We evaluate IntentGC against baselines such as DeepWalk, GraphSage, DSPR, Metapath2vec++, and BiNE on Taobao and Amazon datasets. Offline metrics and online A/B tests on Taobao demonstrate that IntentGC consistently outperforms these methods. Additionally, IntentNet’s efficiency on billion‑scale graphs surpasses GraphSage.

5. Conclusion and Future Work

IntentGC presents a novel, scalable graph convolution framework that effectively incorporates massive unlabeled heterogeneous information for recommendation. Experiments confirm its superiority on large‑scale e‑commerce graphs. Future directions include applying the framework to other tasks and extending it to dynamic graph convolution models that can capture real‑time user features.