SparseNet: Sparse Tweet Network for Classification of Informative Posts using Graph Convolutional Network

A major challenge during critical events is to differentiate useful content from the noise spread through posts on social media. This paper presents the use of a Graph Convolutional Network (GCN) to classify tweets as informative or non-informative on a new crisis dataset. To reduce the computational complexity, we evaluate the performance of the GCN by varying the degree of sparseness of the input tweet graph. Our experimental results show that the classification results improve when the input graph is sparse and contains edges only between highly similar tweets compared to a complete graph. We show the efficacy of our approach by achieving an accuracy of 88.40% and an F1-score of 88.05% despite using a small-sized dataset.