Heterogeneous deep graph convolutional network with iterative deep graph learning for Covid-19 inline recommendation

Covid-19 has been causing persistent negative impact on people’s daily life all over the world. A mass of scientific literature in medicine and biology has been published since then. It could be a hard process to gather the needed papers for researchers and scientists. Many approaches have been proposed to generate inline recommendation list for paper searching. However, the previous methods have some defects. For example, the content-based method only focuses on the full text but neglects the citation relation. Some other deep learning methods did not take the refresh of graph structure into consideration. Hence, this paper presents a new deep learning method for the COVID-19 citation recommendation task. Two main parts are included in this method. The first part, iterative deep graph learning (IDGL), is to optimize the graph structure by iterating the similarity learning process. The second part, heterogeneous deep graph convolution network (HDGCN), is to learn effective dense node vectors. The node embedding vectors are first initialized by BioBERT and put into the heterogeneous graph convolution layer. The heterogenous graph convolution layer learns the graph with nodes of different features. Then the data flows into the deep attention layer, which reweights node importance using the trainable vectors simultaneously. Finally, the data goes through the decoder layer to recover the graph structure and generates the recommendation list based on the similarity score. It is proved that the method presented in this paper performs better than the previous methods on the CORD-19 dataset and Litcovid dataset.