[Paper Review] Subgraph Neural Networks

Emily Alsentzer, Samuel G. Finlayson+2 more|arXiv (Cornell University)|Jun 18, 2020

Advanced Graph Neural NetworksComputer Science84 references44 citations

TL;DR

SubGNN learns disentangled subgraph representations through a novel subgraph-level message-passing framework with three property-aware channels (position, neighborhood, structure) and anchor-patch routing, achieving strong subgraph classification on synthetic and real-world datasets.

ABSTRACT

Deep learning methods for graphs achieve remarkable performance on many node-level and graph-level prediction tasks. However, despite the proliferation of the methods and their success, prevailing Graph Neural Networks (GNNs) neglect subgraphs, rendering subgraph prediction tasks challenging to tackle in many impactful applications. Further, subgraph prediction tasks present several unique challenges: subgraphs can have non-trivial internal topology, but also carry a notion of position and external connectivity information relative to the underlying graph in which they exist. Here, we introduce SubGNN, a subgraph neural network to learn disentangled subgraph representations. We propose a novel subgraph routing mechanism that propagates neural messages between the subgraph's components and randomly sampled anchor patches from the underlying graph, yielding highly accurate subgraph representations. SubGNN specifies three channels, each designed to capture a distinct aspect of subgraph topology, and we provide empirical evidence that the channels encode their intended properties. We design a series of new synthetic and real-world subgraph datasets. Empirical results for subgraph classification on eight datasets show that SubGNN achieves considerable performance gains, outperforming strong baseline methods, including node-level and graph-level GNNs, by 19.8% over the strongest baseline. SubGNN performs exceptionally well on challenging biomedical datasets where subgraphs have complex topology and even comprise multiple disconnected components.

Motivation & Objective

• Motivate subgraph prediction as a distinct task from node- or graph-level prediction.
• Propose a subgraph-focused message passing architecture that captures position, neighborhood, and structure.
• Introduce anchor-patch routing to aggregate information from across the underlying graph.
• Demonstrate empirical gains on synthetic and real-world subgraph classification datasets.

Proposed method

• Define subgraphs S with potential multiple components and labels.
• Introduce E_S: a subgraph-level message passing scheme operating on subgraph components.
• Introduce three channels (Position, Neighborhood, Structure) with dedicated anchor-patch sampling, encoding, and similarity routing.
• Compute property-aware outputs z_x,c per channel and aggregate them into the final subgraph embedding z_S.
• Provide computationally efficient architecture with fixed anchor patches and optional precomputation of similarities.
• Demonstrate extension potential to unsupervised, semi-supervised, or other prediction tasks involving subgraphs.

Experimental results

Research questions

• RQ1 How can subgraphs be represented to capture their internal topology, border structure, and relative position within the host graph?
• RQ2 Can a subgraph-focused GNN with channel-specific routing outperform node- or graph-level baselines on subgraph classification tasks?
• RQ3 Do anchor patches and property-aware routing effectively disentangle subgraph properties (position, neighborhood, structure) for accurate predictions?
• RQ4 How do SubGNN variants perform on synthetic benchmarks designed for subgraph topology versus real-world biomedical and social datasets?

Key findings

• SubGNN outperforms baselines by 77.4% on synthetic datasets and by 125.2% on real-world datasets on average.
• SubGNN improves the strongest baseline by 19.8% on overall subgraph classification tasks.
• Channel ablation shows each channel’s performance aligns with its intended inductive bias (e.g., structure channel excels on border-structure tasks, position channel on internal-component relationships).
• SubGNN achieves strong results on challenging biomedical datasets where subgraphs have complex topology or multiple components.