WikipediaThe free encyclopedia

arXiv Semantic Graph

1 language
  • Article
  • Read
From Wikipedia, the free encyclopedia
This article is about the interactive demo. For the search engine that hosts it, see David's Internet.
arXiv Semantic Graph
TypeInteractive demo
OriginDTU course 02807, Computational Tools for Data Science
DeveloperDavid
Written inPython, TensorFlow Hub, NetworkX, TypeScript
TestsLouvain and from-scratch algorithms run live on a 2,500-paper subsample
Built withDTU Group 36 project, replayed on a stratified subsample
Hosted onDavid-Internet/demos/arxiv
Websitehttps://david-internet.vercel.app/demos/arxiv

arXiv Semantic Graph is an interactive demonstration of a recommendation-system project by Group 36 in DTU course 02807, Computational Tools for Data Science. It is hosted on David's Internet and displays a 2,500-paper stratified subsample.[1]

The project transforms arXiv abstracts into Universal Sentence Encoder embeddings, joins similar papers into a graph, finds Louvain communities, and supplies nearest-neighbour recommendations. It also includes live implementations of A-priori, Girvan-Newman, and Laplacian spectral clustering.

Overview

Visitors can drag a similarity threshold, select a paper, and inspect its five nearest neighbours with same-community labels. The course algorithm panels use 14,963 grocery baskets for A-priori, Zachary's karate club for Girvan-Newman, and a Jacobi eigensolver for the Laplacian view.

Semantic graph

The original pipeline used 148,477 abstracts, HNSW nearest-neighbour search, and a global distance threshold τ. The archived conclusion was that modularity peaked at tau 0.19 and the team shipped 0.27. The first value maximized modularity in the full run, while the second was selected in the report for usability. The browser graph uses exact brute-force nearest neighbours on its smaller subsample.

Development

The demo replays a fall 2025 Group 36 project. It contains fixture-tested TypeScript ports of Louvain and modularity, threshold selection, A-priori, Girvan-Newman, and a Jacobi eigensolver.Louvain and from-scratch algorithms run live on a 2,500-paper subsample.

See also

References

  1. ^ content/arxiv/README.md, David's Internet.
Categories:
  • Interactive demos
  • Information retrieval
  • 2026 establishments