Recurrent Convolution based Graph Neural Network for Node Classification in Graph structure data

Many datasets in various machine learning applications are structural and naturally represented as graphs. They comprise data from the analyses of social and communication networks, predictions of traffic, and fraud detection. Graphbased Deep Learning (DL) aims to construct and train graph datasets attuned models for various graph-structured based tasks. In this work, we presented a model of Graph Neural Network (GNN) for the node classification task. We have compared our proposed model with a baseline model on three citation network datasets: CORA, PUBMED, and CITESEER. We examined the baseline and proposed models predictions on new data instances by randomly generating binary work vectors concerning the work presence probabilities for all three datasets. The proposed model is significantly better than the baseline model on the CORA and CITESEER datasets.

I. Introduction

Node Classification is a basic task in graph-structure data. A graph-structure data comprises nodes (vertices) and links (edges). A simple undirected graph ‘G’ is described by , in which represents the set of nodes, and ‘E’ is the set of relationships or links between the nodes. Many real-world data are naturally represented in a graph, such as, in a recommender system, products and customers can be considered nodes or vertices. The relationships between the purchased product and the customer are the edges or links. By using a graph, we can draw the spending habits of customers. Furthermore, a node of a graph can have attributes of features. For example, a customer has features like gender and age and product features like size and price. Other system like Social Networks [1], Physical Network [2], Traffic [3], Natural Language Processing [4], and so on are also represented as huge graphs.