Periodical embeddings uncover hidden interdisciplinary patterns in the subject classification scheme of science

TL;DR

This work tackles the challenge of evaluating journal classification schemes by proposing Periodical2Vec (P2V), a data-driven embedding learned from periodical-level citations. Using a massive corpus of over $23{,}322{,}430$ abstracts, the authors show that k-means clustering on P2V embeddings yields superior semantic coherence and predictive performance compared to Scopus and other citation-based baselines. Through document classification, topic-label consistency with LDA, and structural analyses (Sankey flows and 2-D maps), the study reveals both refined disciplinary boundaries and coherent interdisciplinary clusters, underscoring the dynamic, interconnected nature of modern science. The findings suggest that citation-derived periodical embeddings offer a robust, scalable, and interpretable map of science with practical implications for evaluation, funding, and science policy, while also outlining avenues for multimodal extensions to further enhance semantic capture.

Abstract

Subject classification schemes are foundational to the organization, evaluation, and navigation of scientific knowledge. While expert-curated systems like Scopus provide widely used taxonomies, they often suffer from coarse granularity, subjectivity, and limited adaptability to emerging interdisciplinary fields. Data-driven alternatives based on citation networks show promise but lack rigorous, external validation against the semantic content of scientific literature. Here, we propose a novel quantitative framework that leverages classification tasks to evaluate the effectiveness of journal classification schemes. Using over 23 million paper abstracts, we demonstrate that labels derived from k-means clustering on Periodical2Vec (P2V)--a periodical embedding learned from paper-level citations--yield significantly higher classification performance than both Scopus and other data-driven baselines (e.g., citation, co-citation, and Node2Vec variants). By comparing journal partitions across classification schemes, two structural patterns emerge on the map of science: (1) the reorganization of disciplinary boundaries--splitting overly broad categories (e.g., "Medicine" into "Oncology", "Cardiology", and other specialties) while merging artificially fragmented ones (e.g., "Chemistry" and "Chemical Engineering"); and (2) the identification of coherent interdisciplinary clusters--such as "Biomedical Engineering", "Medical Ethics", and "Information Management"--that are dispersed across multiple categories but unified in citation space. These findings underscore that citation-derived periodical embeddings not only outperform traditional taxonomies in predictive validity but also offer a dynamic, fine-grained map of science that better reflects both the specialization and interdisciplinarity inherent in contemporary research.

Figures (3)

• Figure 1: One-vs-rest classification performance comparison between classifiers trained using different classification labels. Curves are plotted as the macro average of all individual classes' results. (A) Precision-Recall curves. The number in parentheses indicates average precision. (B) Receiver operating characteristic curves. The number in parentheses indicates the area under the curve.
• Figure 2: A Sankey diagram visualizing how journals are assigned differently between the Scopus categories (left-side nodes) and the data-driven clusters (right-side nodes). The top 5 frequent words from journal names are presented for each node on the right side to indicate cluster topics. The thickness of flows quantifies the proportion of journals redistributed between the two classification schemes. For visual clarity, we filter out flows with less than $\min(10\%\times N_{source},50)$ journals, where $N_{source}$ is the number of journals in source nodes (the left side). An interactive online version of this Sankey diagram can be accessed at https://lyuzhuoqi.github.io/periodical-clustering/sankey/snakey_kmeans_filtered.html.
• Figure 3: Comparing $\mathcal{C}^{Scopus}$ and $\mathcal{C}^{P2V}$. (A--B) 2-D projections of journal P2V embeddings using t-SNE van2008visualizing. Legends are shared between (A) and (B) to enable visual cluster tracking. (A) Partitioned by $\mathcal{C}^{Scopus}$. Color denotes label. (B) Partitioned by $\mathcal{C}^{P2V}$. Color denotes cluster #. (C) A interpolated heatmap of element-wise clustering similarity between Scopus and clustering-generated label, with exemplary topic-cohesive clusters which have notable low/high similarity annotated (check detailed cluster members in Tables S2--S5). The line types of arrows indicate the reason for the unusual similarity between 2 classification schemes. The interpolation is based on inverse distance weighting (IDW) using power parameter=2.