A Fine-grained Academic ViewPoints Network Dataset for Graph Structure Analysis (VPN)

This article presents the ViewPoints Network (VPN) dataset, a fine-grained academic graph resource designed to support the analysis of viewpoint-level structures within scientific literature. While existing academic graph datasets primarily operate at the paper or author level, they often capture only coarse relationships such as citations or co-authorship, leaving the semantic evolution of scientific ideas insufficiently represented. The VPN dataset addresses this gap by extracting $1,474,680$ viewpoints from each publication abstract and representing them as individual nodes with high-dimensional sentence embeddings. Similarity-based edges are constructed using multiple thresholds, enabling flexible formation of viewpoint graphs tailored to different analytical needs. The dataset covers eight research domains in computer science and provides structured metadata, embedding files, and multi-threshold adjacency matrices to facilitate diverse graph learning and semantic analysis tasks. All data are processed through a transparent pipeline including cleaning, viewpoint extraction using large language models, embedding generation, and similarity computation.By offering a more granular representation of academic discourse, the VPN dataset supports research on viewpoint clustering, semantic interaction, idea evolution, and fine-grained network analysis across scientific domains.