Knowledge Prediction by Graph Embedding and Machine Learning

Abstract

This paper is an extended research of the project “The Knowledge Database/ Graph of China-studies” (https://reurl.cc/35Ak59) . The main research target is to predict the new research stream from known journal papers by the graph embedding and link prediction. The challenge of our dataset does not include citation relationships; therefore, we might retrieve features of relationships from the content of the papers inside directly. We used keywords collaboration and k-means to reduce dimension, then word2vec and MLP to classify if any two nodes can link in the next round (year). Finally, we could achieve over 90% accuracy in each round which is better than the base-line method (random-forest with Adar and Jaccard score). And we also provide a visualization graph in action. We contribute a pipeline workflow to the rawer bibliography dataset which doesn’t conclude cite-relationship, and this workflow can be used on social media or other text-only datasets.