Eponymous Author Disambiguation Method Based on Multi-scale and Clustering Properties in Graph Neural Networks

Fang yuan, DengHao, Yangxiangwu

Keywords: Graph Neural Networks, Multi-Scale, Clustering, Feature-Enhanced Tree Model

TL;DR: Eponymous Author Disambiguation Method Based on Multi-scale and Clustering Properties in Graph Neural Networks

Abstract:

The WhoIsWho-IND dataset challenge requires developing a model to identify incorrectly assigned papers for each author profile, including name and publications. The KDD CUP OAG Challenge involves authenticating papers from a large dataset, combining clustering and classification without prior research. This paper introduce that a graph neural network leveraging multi-scale and clustering to extract and splice diverse features, which are processed through a clustering layer and then classified. Trained on 515 authors, it achieved an AUC of 0.78 on a 370-author test set, outperforming a feature-enhanced tree model by 3% accuracy.

Submission Number: 16