Large-Scale Learnable Graph Convolutional Networks

Abstract

Convolutional neural networks (CNNs) have achieved great success on grid-like data such as images, but face tremendous challenges in learning from more generic data such as graphs. In CNNs, the trainable local filters enable the automatic extraction of high-level features. The computation with filters requires a fixed number of ordered units in the receptive fields. However, the number of neighboring units is neither fixed nor are they ordered in generic graphs, thereby hindering the applications of convolutional operations. Here, we address these challenges by proposing the learnable graph convolutional layer (LGCL).LGCL automatically selects a fixed number of neighboring nodes for each feature based on value ranking in order to transform graph data into grid-like structures in 1-D format, thereby enabling the use of regular convolutional operations on generic graphs. To enable model training on largescale graphs, we propose a sub-graph training method to reduce the excessive memory and computational resource requirements suffered by prior methods on graph convolutions. Our experimental results on node classification tasks in both transductive and inductive learning settings demonstrate that our methods can achieve consistently better performance on the Cora, Citeseer, Pubmed citation network, and protein-protein interaction network datasets. Our results also indicate that the proposed methods using subgraph training strategy are more efficient as compared to prior approaches.

Attributed Multi-Relational Attention Network for Fact-checking URL Recommendation

“…The core idea is to iteratively aggregate attributed node vectors around each node, and messages propagates by stacking multiple layers. However, the original design of GCN is not suitable for our scenario because of the following reasons: First, existing GCN works [12,14] do not distinguish different types of nodes, whereas in our case, it does not make sense to aggregate user and URL nodes together. And the aggregation function proposed in most GCN works treats all its adjacency nodes with the same importance.…”

“…With the surge of Graph-based Neural Network, GCN-based approaches have shown strong effectiveness on various tasks [12,14,23], including recommender system. The core idea is to iteratively aggregate attributed node vectors around each node, and messages propagates by stacking multiple layers.…”

GraphDTA: Predicting drug–target binding affinity with graph neural networks

“…Compared to the spectralbased methods which handle the whole graph simultaneously, the spatial approaches can instead process graph nodes in batches thus can be scalable to large graphs. Recent works on this approach include [21,23,12,37,36,40].…”

“…Compared to the spectral-based methods which handle the whole graph simultaneously, the spatial approaches can instead process graph nodes in batches thus can be scalable to large graphs. Recent works on this approach include [21,23,12,37,36,40] GCNs have been used in computational drug discovery [34], including quantitative structure activity/property relationship prediction, interaction prediction, synthesis prediction, and de novo molecular design. The problem we explore in this paper, prediction of drug-target binding affinity, belongs to the task of interaction prediction, where the interactions could be among drugs, among proteins, or between drugs and proteins.…”

Graph-Based Method for Anomaly Detection in Functional Brain Network using Variational Autoencoder

“…Graph Convolutional Neural Network (G-CNN): for apply convolution-like operators over irregular local supports, as graphs where nodes can have a varying number of neighbors which can be used as layers in deep networks, for node classification or recommendation, link prediction and etc. in this process we involved with three challenges, a) defining translation structure on graphs to allow parameters sharing, b) designing compactly supported filters on graphs, c) aggregating multi-scale information, the proposed strategies broadly fall into two domains, there is one spatial operation directly perform the convolution by aggregating the neighbor nodes' information in a certain batch of the graph, where weights can be easily shared across different structures [21,22] and other one is spectral operation relies on the Eigen-decomposition of the Laplacian matrix that apply in whole graph at the same time [23,24,25,26], spectral-based decomposition is often unstable making the generalization across different graphs difficult [10], that cannot preserve both the local and global network structures also require large memory and computation. On the other hand, local filtering approaches [27] rely on possibly suboptimal hardcoded local pseudo-coordinates over graph to define filters.…”