[论文解读] Supervised Community Detection with Line Graph Neural Networks

Zhengdao Chen, Xiang Li+1 more|arXiv (Cornell University)|May 23, 2017

Complex Network Analysis Techniques参考文献 43被引用 120

一句话总结

The paper introduces Line Graph Neural Networks (LGNNs) that augment GNNs with the line graph and non-backtracking operator to improve supervised community detection, achieving or surpassing belief propagation in several SBM regimes and performing well on real datasets.

ABSTRACT

Traditionally, community detection in graphs can be solved using spectral methods or posterior inference under probabilistic graphical models. Focusing on random graph families such as the stochastic block model, recent research has unified both approaches and identified both statistical and computational detection thresholds in terms of the signal-to-noise ratio. By recasting community detection as a node-wise classification problem on graphs, we can also study it from a learning perspective. We present a novel family of Graph Neural Networks (GNNs) for solving community detection problems in a supervised learning setting. We show that, in a data-driven manner and without access to the underlying generative models, they can match or even surpass the performance of the belief propagation algorithm on binary and multi-class stochastic block models, which is believed to reach the computational threshold. In particular, we propose to augment GNNs with the non-backtracking operator defined on the line graph of edge adjacencies. Our models also achieve good performance on real-world datasets. In addition, we perform the first analysis of the optimization landscape of training linear GNNs for community detection problems, demonstrating that under certain simplifications and assumptions, the loss values at local and global minima are not far apart.

研究动机与目标

• Motivate community detection as a supervised node-classification task on graphs across distributions of input graphs.
• Develop GNN architectures that exploit multiscale graph operators and the line graph to capture higher-order edge interactions.
• Demonstrate data-driven performance gains over traditional spectral methods and belief propagation on SBM and GBM models.
• Provide an analysis of the optimization landscape for linear GNNs in community detection.
• Show applicability to real-world networks from SNAP and discuss learned representations.
• Assess the impact of edge-oriented information flow via the non-backtracking operator on detection performance.

提出的方法

• Define a multiscale GNN layer using a family of graph operators {I, D, A, AJ} for node features.
• Augment GNNs with the line graph and non-backtracking operator to create LGNN, enabling directed edge information flow.
• Construct LGNN with interactions between node and edge states via incidence matrices, and allow communication between nodes and edges each layer.
• Define a permutation-invariant loss over community labels to account for label symmetry.
• Train end-to-end with backpropagation on graphs of varying sizes, using instance normalization for stability.
• Compare GNN and LGNN variants (including linear and symmetric LGNNs) against BP, spectral methods, and GAT on SBM, GBM, and SNAP datasets.

实验结果

研究问题

• RQ1 Can supervised GNNs approximate or surpass belief propagation for community detection on sparse graphs?
• RQ2 Does incorporating line-graph based, non-backtracking information improve detection in SBM/GBM regimes, including hard computational-to-statistical gaps?
• RQ3 How do GNN-based methods perform on real datasets with overlapping and unbalanced communities compared to traditional models?
• RQ4 What is the role of linear (non-activated) GNNs in achieving competitive performance and what does their optimization landscape look like?
• RQ5 How does a permutation-invariant loss affect learning in community detection tasks across graphs of varying sizes?

主要发现

Model	Avg Overlap	Std Dev
GNN	0.18	0.04
LGNN	0.21	0.05
LGNN-L	0.18	0.04
LGNN-S	0.18	0.04
GAT	0.16	0.04
BP	0.14	0.02

• GNNs and LGNNs match or surpass belief propagation on binary SBMs and multi-class SBMs within hard regimes.
• Linear LGNNs achieve performance close to BP, consistent with spectral approximations via Bethe Hessian.
• LGNN with non-backtracking edge information outperforms symmetric LGNN and GAT in several SBM experiments.
• LGNNs retain strong performance on GBM and real SNAP datasets, indicating practical applicability beyond synthetic models.
• The learning landscape of linear GNNs is benign under certain assumptions, with local minima near global minima as graph size grows.