Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiGNN: Graph Neural Networks at LinkedIn

Fedor Borisyuk, Shihai He, Yunbo Ouyang, Morteza Ramezani, Peng Du, Xiaochen Hou, Chengming Jiang, Nitin Pasumarthy, Priya Bannur, Birjodh Tiwana, Ping Liu, Siddharth Dangi, Daqi Sun, Zhoutao Pei, Xiao Shi, Sirou Zhu, Qianqi Shen, Kuang-Hsuan Lee, David Stein, Baolei Li, Haichao Wei, Amol Ghoting, Souvik Ghosh

TL;DR

LiGNN presents a production-ready, large-scale GNN framework at LinkedIn that tackles scale, heterogeneity, cold-start, and temporal dynamics by integrating a heterogeneous graph with a DeepGNN-based training pipeline, temporal sequence modeling, graph densification, and near-line inference. The approach combines an encoder–decoder architecture with GraphSAGE-style inductive learning, multiple sampling strategies (notably 2-hop PPR), and production-aware optimizations (adaptive neighbor sampling, grouping/slicing, and shared-memory queues) to achieve substantial online gains across Follow Feed, Job, People, and Ads, while reducing training time from days to hours. Deployment lessons emphasize real-time compute graphs via GE, impression-discount ordering, and system-level stability to enable reliable production use. Overall, LiGNN delivers practical, scalable solutions for applying GNNs at large scale with measurable business impact.

Abstract

In this paper, we present LiGNN, a deployed large-scale Graph Neural Networks (GNNs) Framework. We share our insight on developing and deployment of GNNs at large scale at LinkedIn. We present a set of algorithmic improvements to the quality of GNN representation learning including temporal graph architectures with long term losses, effective cold start solutions via graph densification, ID embeddings and multi-hop neighbor sampling. We explain how we built and sped up by 7x our large-scale training on LinkedIn graphs with adaptive sampling of neighbors, grouping and slicing of training data batches, specialized shared-memory queue and local gradient optimization. We summarize our deployment lessons and learnings gathered from A/B test experiments. The techniques presented in this work have contributed to an approximate relative improvements of 1% of Job application hearing back rate, 2% Ads CTR lift, 0.5% of Feed engaged daily active users, 0.2% session lift and 0.1% weekly active user lift from people recommendation. We believe that this work can provide practical solutions and insights for engineers who are interested in applying Graph neural networks at large scale.

LiGNN: Graph Neural Networks at LinkedIn

TL;DR

LiGNN presents a production-ready, large-scale GNN framework at LinkedIn that tackles scale, heterogeneity, cold-start, and temporal dynamics by integrating a heterogeneous graph with a DeepGNN-based training pipeline, temporal sequence modeling, graph densification, and near-line inference. The approach combines an encoder–decoder architecture with GraphSAGE-style inductive learning, multiple sampling strategies (notably 2-hop PPR), and production-aware optimizations (adaptive neighbor sampling, grouping/slicing, and shared-memory queues) to achieve substantial online gains across Follow Feed, Job, People, and Ads, while reducing training time from days to hours. Deployment lessons emphasize real-time compute graphs via GE, impression-discount ordering, and system-level stability to enable reliable production use. Overall, LiGNN delivers practical, scalable solutions for applying GNNs at large scale with measurable business impact.

Abstract

In this paper, we present LiGNN, a deployed large-scale Graph Neural Networks (GNNs) Framework. We share our insight on developing and deployment of GNNs at large scale at LinkedIn. We present a set of algorithmic improvements to the quality of GNN representation learning including temporal graph architectures with long term losses, effective cold start solutions via graph densification, ID embeddings and multi-hop neighbor sampling. We explain how we built and sped up by 7x our large-scale training on LinkedIn graphs with adaptive sampling of neighbors, grouping and slicing of training data batches, specialized shared-memory queue and local gradient optimization. We summarize our deployment lessons and learnings gathered from A/B test experiments. The techniques presented in this work have contributed to an approximate relative improvements of 1% of Job application hearing back rate, 2% Ads CTR lift, 0.5% of Feed engaged daily active users, 0.2% session lift and 0.1% weekly active user lift from people recommendation. We believe that this work can provide practical solutions and insights for engineers who are interested in applying Graph neural networks at large scale.
Paper Structure (26 sections, 1 equation, 7 figures, 8 tables, 3 algorithms)

This paper contains 26 sections, 1 equation, 7 figures, 8 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. GNN Modeling and Training
  4. Graph Construction
  5. Training infra
  6. GNN architecture
  7. Temporal Graphs
  8. Graph Densification
  9. Multi-hop Graph Sampling
  10. Training Stability and Speed
  11. Training stability
  12. Training speed
  13. Near-line inference
  14. Experiments
  15. Experiments in Follow Feed
  16. ...and 11 more sections

Figures (7)

  • Figure 1: Schematic representation of LinkedIn Graph. Members engaging with Posts, Jobs, Groups, Companies and other members.
  • Figure 2: High level view of GNN pipelines.
  • Figure 3: GNN model architecture.
  • Figure 4: Temporal model includes: (1) static SAGE-encoder depicted in Yellow, (2) transformer based temporal sequence model in Green. Red lines depict long term loss and blue lines depict cosine similarity loss described in §\ref{['sec:gnn_architecture']}.
  • Figure 5: Follow Feed with Adaptive Neighbor Sampling and Grouping & Slicing
  • ...and 2 more figures