Author Name Disambiguation Based on Multi Relationship Fusion and Representation Learning

Author Name disambiguation constitutes a critical technology for building scholar profiles and enabling accurate academic search systems. While existing unsupervised name disambiguation approaches often rely on simplistic combinations of explicit paper relationships while overlooking latent network structures, this study develops an advanced disambiguation framework through multi-relational feature fusion. Our model systematically constructs a scholarly information network incorporating both semantic features (professional term sets from titles, abstracts and keywords) and relational patterns (implicit connections via important authors), learns unified paper representations using Variational Graph Autoencoder (VGAE), and produces final clusters through Hierarchical Agglomerative Clustering (HAC). Experimental validation on the AMiner-na benchmark demonstrates significant performance gains, with ournum and our-distance methods achieving F1-scores of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 8. 9 \%}$</tex> and 63.25 % respectively, substantially outperforming conventional baselines by effectively exploiting both explicit and implicit academic data characteristics.