Graph Tree Neural Networks

Seokjun Kim, Jaeeun Jang, Heeseok Jung, Hyeoncheol Kim

Keywords: Graph neural networks, Tree based convolutional neural network, Domain general learning

Abstract: In the field of deep learning, various architectures have been developed. However, most studies are limited to specific tasks or datasets due to their fixed layer structure. In this paper, we do not express the structure delivering information as a network model but as a data structure called a graph tree. And we propose two association models of graph tree neural networks(GTNNs) designed to solve the problems of existing networks by analyzing the structure of human neural networks. Defining the starting and ending points in a single graph is difficult, and a tree cannot express the relationship among sibling nodes. On the contrary, a graph tree(GT) can express leaf and root nodes as its starting and ending points and the relationship among sibling nodes. Instead of using fixed sequence layers, we create a GT for each data and train GTNN according to the tree's structure. GTNNs are data-driven learning in which the number of convolutions varies according to the depth of the tree. Moreover, these models can simultaneously learn various types of datasets through the recursive learning method. 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. To demonstrate the performance of these networks, we conducted two experiments. The first experiment is whether various datasets can be processed by combining GTNN and feature extraction networks(processing various datasets). The second experiment is about whether the output of GTNN can embed information on all data contained in the GT(association). We compared the performance of existing networks that separately learned image, sound, and natural language datasets with the performance simultaneously learned by connecting these networks. As a result, these models learned without significant performance degradation, and the output vector contained all the information in the GT.

One-sentence Summary: We propose graph tree neural networks that is freer from architecture by integrating various developed models.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2111.00424/code)