Abstract
Graph neural networks (GNNs) have recently shown good performance in various fields. In this paper, we propose graph tree neural networks (GTNNs) designed to solve the problems of existing networks by analyzing the structure of human neural networks. In GTNNs, information units are related to the form of a graph and then they become a bigger unit of information again and have a relationship with other information units. At this point, the unit of information is a set of neurons, and we can express it as a vector with GTNN. Defining the starting and ending points in a single graph is difficult, and a tree cannot express the relationship among sibling nodes. However, a graph tree can be expressed using leaf and root nodes as its starting and ending points and the relationship among sibling nodes. Depth-first convolution (DFC) encodes the interaction result from leaf nodes to the root node in a bottom-up approach, and depth-first deconvolution (DFD) decodes the interaction result from the root node to the leaf nodes in a top-down approach. GTNN is data-driven learning in which the number of convolutions varies according to the depth of the tree. Moreover, learning features of different types together is possible. Supervised, unsupervised, and semi-supervised learning using graph tree recursive neural network (GTR) , graph tree recursive attention networks (GTRAs), and graph tree recursive autoencoders (GTRAEs) are introduced in this paper. We experimented with a simple toy test with source code dataset.
Abstract (translated)
URL
https://arxiv.org/abs/2111.00424
