Text Classification Based on GNN

Abstract

The phenomenon that AI researchers tend to transform some certain data into the form of graphs is prevailing. Usually, these graph-like data will be inputted into some certain artificial neutral networks which are dramatically disparate with the conventional CNN. The purpose of the algorithm employed the GNN is to extract much more detailed features that can be stored easily in the graph. However, these detailed features are much more difficult to extract in the raw data accumulated in the database, which requires the experiment to transfer the common database into the whole graph ahead of schedule. The dataset used in this paper, Cora, is commonly used in some papers whose targets aimed at semantic segmentation, while disparate with this paper as well. The result of this experiment has achieved to nearly 100% accuracy accompanied with those preprocessed data. Furthermore, this paper also attaches much focus on the effects of preprocessing operation which can be reflected on the differences of accuracy. Only by preprocessing operation can this paper achieve better results accompanied with higher accuracy when compared with other experiments.