Graph-based entity resolution and completion for academic knowledge graphs

Abstract

[en] This thesis explores graph-based entity resolution and completion within academic knowledge graphs, focusing on the complex relationships between authors and papers and between papers themselves using Graph Neural Networks (GNNs). Raw data sourced from the American Physical Society underwent meticulous data cleaning and entity resolution analysis to prepare it for the proposed network. Author grouping strategies and citation overlap were examined, revealing distinct clusters of researchers and insightful patterns in citation relationships. A GNN model was developed using SAGEConv layers and heterogeneous transformations to capture local graph structures for accurate link prediction. This model was optimized with mini-batch loading and edge-level splits, which contributed to its high accuracy in predicting links between authors and papers, as demonstrated in the evaluation. The findings underscore the model’s capability to uncover hidden relationships and trends within the academic graph. Future work could enhance the model by incorporating additional features, experimenting with alternative GNN architectures, and including more detailed citation contexts and collaboration networks. Overall, this thesis highlights the transformative potential of GNNs in entity resolution and completion for academic knowledge graphs.