Research On Academic Influence Prediction Algorithm Based On Representation Learning

Scientific impact prediction is an important research content of scholarly data application.Paper impact prediction can help researchers find valuable papers to read and follow the latest research direction.Author impact prediction helps to provide reference for decision makers in fund funding and resource allocation.The most advanced methods of predicting academic influence are graph-based approaches.One limitation of the graph-based approaches is to focus on the current importance of the literature and researchers,which is unfair to newly published articles and young researchers,another limitation is that they rely primarily on the global structural characteristics of academic network,which cannot fully utilize the local structural information of the network and other rich information,such as text and time information.To address these problems,the paper proposes a representation learning-based scientific impact prediction model,which mainly includes the following contents:A heterogeneous academic network which contains various types of relationships between papers,authors,topics and words is constructed,and the weights of the edges are set according to the characteristics related to the impact of the paper and the author.The content relevance between papers is captured through topics and words,and then the potential impact of newly published papers which lack citations can be fully exploited.A representation learning model is designed for heterogeneous academic network to learn the representations of different types of nodes.The heterogeneous academic network is mapped into a low-dimensional latent space by constructing appropriate objective functions.The local structural information of the network and the text information of papers can be effectively preserved through the process of representing learning.The similarities between the representations of nodes are used to construct the adjacency matrices corresponding to the paper citation network,the paper-author network and the Co-author network.Then a graph-based multivariate random walk model is designed based on the mutual reinforcement relationship between papers and authors,to rank the future impact of the paper and author simultaneously.Extensive experiments results on the AMiner dataset and the AAN dataset show that the proposed method significantly outperforms existing five approaches for the two evaluation metrics of Normalized Discounted Cumulative Gain and Recommended Intensity.