[Paper Review] Search-Convolutional Neural Networks

James Atwood, Don Towsley|arXiv (Cornell University)|Nov 6, 2015

Advanced Graph Neural Networks19 references20 citations

TL;DR

This paper introduces Search-Convolutional Neural Networks (SCNNs), a deterministic relational model that extends convolution to graph search to extract local structural patterns from graph-structured data. SCNNs achieve size-independent parameterization, enabling transfer learning and outperforming standard classifiers and multilayer perceptrons while matching state-of-the-art probabilistic models at lower computational cost.

ABSTRACT

We present a new deterministic relational model derived from convolutional neural networks. Search-Convolutional Neural Networks (SCNNs) extend the notion of convolution to graph search to construct a rich latent representation that extracts local behavior from general graph-structured data. Unlike other neural network models that take graph-structured data as input, SCNNs have a parameterization that is independent of input size, a property that enables transfer learning between datasets. SCNNs can be applied to a wide variety of prediction tasks, including node classification, community detection, and link prediction. Our results indicate that SCNNs can offer considerable lift over off-the-shelf classifiers and simple multilayer perceptrons, and comparable performance to state-of-the-art probabilistic graphical models at considerably lower computational cost.

Motivation & Objective

• To develop a neural network model that effectively captures local structural patterns in graph-structured data.
• To design a parameterization independent of input size to enable transfer learning across datasets.
• To improve performance on graph prediction tasks such as node classification, community detection, and link prediction.
• To achieve competitive performance with state-of-the-art probabilistic graphical models at significantly reduced computational cost.

Proposed method

• SCNNs apply a convolution-like operation based on graph search, replacing traditional spatial convolution with a search-driven aggregation over local graph neighborhoods.
• The model constructs a latent representation by traversing local subgraphs using a predefined search strategy, capturing rich relational inductive biases.
• Parameter sharing is achieved through a fixed-size, search-based receptive field, ensuring parameterization independence from input graph size.
• The architecture supports end-to-end training for various prediction tasks using standard backpropagation.
• The method generalizes across different graph types and scales efficiently due to its size-invariant design.

Experimental results

Research questions

• RQ1 Can a deterministic neural network model effectively extract local structural patterns from general graph-structured data?
• RQ2 Does a search-based convolution mechanism outperform standard convolution or fully connected layers on graph prediction tasks?
• RQ3 Can a size-independent parameterization enable effective transfer learning across diverse graph datasets?
• RQ4 How does SCNN performance compare to state-of-the-art probabilistic graphical models in terms of accuracy and computational efficiency?

Key findings

• SCNNs achieve significant performance gains over off-the-shelf classifiers and multilayer perceptrons on graph prediction tasks.
• SCNNs match the performance of state-of-the-art probabilistic graphical models while requiring considerably less computational cost.
• The size-independent parameterization enables effective transfer learning between datasets with different graph sizes.
• SCNNs demonstrate strong generalization across diverse tasks, including node classification, community detection, and link prediction.