A Shallow Graph Neural Network with Innovative Node Updating for Online Handwritten Stroke Classification

Stroke classification is important to the layout analysis of online handwritten documents. Due to the diversity of writing styles and the complexity of layout structure, stroke classification is challenging. Graph neural networks (GNNs) is one of the most effective frameworks for stroke classification. However, GNNs has the problem of node over-compression caused by the deep structure of GNNs, which will lead to loss of node information and hence may deteriorate the performance of stroke classification. In this paper, we propose a shallow graph neural network model that is capable of retaining long-term receptive field by constructing a more reasonable graph through edge classification before the node classification step. Moreover, a novel node learning method is used to integrate edge features into nodes, where edge features not only participate in the calculation of node attention weight as in previous GNN based methods, but also participate in the final node integration. Experiments on the IAMonDo dataset show that our proposed method achieves an accuracy of 97.71% that is superior to existing state-of-the-art methods, demonstrating the effectiveness of the proposed method.