Algorithm Co-occurrence Networks: Mapping Academic Influence

Exploring Academic Influence of Algorithms by Co-occurrence Network Based on Full-text of Academic Papers, by Yuzhuo Wang and five coauthors, constructs large-scale algorithm co-occurrence networks from the full text of natural language processing papers. The paper, submitted to arXiv on 23 Jun 2026 (arXiv:2606.24099), covers more than four decades of academic publications and uses deep learning to extract algorithm entities.

How did the authors build the networks?

They used deep learning models to extract algorithm entities from full-text NLP papers, then built three network variants: overall, cumulative, and annual co-occurrence networks. The extraction pipeline and network construction are applied across the corpus so the study can measure algorithm connections at field, temporal, and yearly scales, relying explicitly on full-text mentions rather than title or keyword signals alone.

The authors emphasize full-text analysis to capture richer co-occurrence patterns than metadata-only approaches. The paper frames these co-occurrence networks as a way to study collective influence formed through algorithm interconnections rather than evaluating algorithms in isolation.

What did they find about algorithm influence?

The networks display typical features of complex networks, and the authors report increasingly dense connections developing over approximately two decades. Classic, high-performing algorithms and those situated at intersections of different research periods tend to score high on popularity, control, centrality, and balanced influence; when an algorithm's influence declines it usually loses its core network position first, then its weaker associations with other algorithms.

The study applies multiple centrality measures to assess group influence across the whole field and over time, comparing overall, cumulative, and annual networks to track shifts. The authors describe the work as "the first large-scale analysis of algorithm co-occurrence networks," and present temporal and structural views that show how algorithm relationships evolve rather than treating each method as an isolated datapoint.

Why does this matter?

A network perspective changes what counts as influence: an algorithm can be central because it connects subfields or because it has many co-mentions, not just because it appears in top-performing results. By extracting entities from full text and measuring connections annually and cumulatively, the study surfaces how algorithms gain or lose structural positions in the field over time, which affects how scholars and evaluators interpret prominence and control.

The paper also provides a methodological foundation for comparative studies: network structural metrics and time-resolved co-occurrence graphs give a repeatable way to compare algorithms, map interdisciplinary intersections, and track historical shifts in NLP research practices.

What to watch

Look for follow-up work that applies the same full-text co-occurrence method beyond NLP or that links these algorithm networks to authors and tasks, since the paper explicitly positions itself as a foundation for future research on networks linking algorithms, scholars, and tasks. The arXiv record lists a journal reference to aslib JIM, 2025, and the submission date on arXiv is 23 Jun 2026 (arXiv:2606.24099), so citation and extension activity in 2026 and beyond will show how the approach is adopted.

References and source notes

• Paper title and abstract as submitted to arXiv:2606.24099, submitted 23 Jun 2026.
• Authors: Yuzhuo Wang, Chengzhi Zhang, Min Song, Seong Deok Kim, Youngsoo Ko, Juhee Lee.
• Journal reference listed in the arXiv record: aslib JIM, 2025.
• Key quoted phrase from the abstract: "the first large-scale analysis of algorithm co-occurrence networks."