CiteTag: Link-Aware Classification for Patents and Manuscripts
An auto-classification service for patent offices and publishers that labels incoming documents using both their text and their citation structure, via a cheap linear model over multi-hop diffusion features — capturing what a document cites and who cites it back.
Concept
Patent offices and publishers must slot every incoming document into a taxonomy — CPC codes for patents, subject areas for journals — mostly with text-only models plus manual review. The strongest cheap signal they ignore is structure: what a document cites and who cites it back. CiteTag builds the citation graph, runs the modified Markov diffusion kernel over K-hop citation neighborhoods, and trains a linear classifier on the resulting per-document features. The kernel's low-pass band carries field-level (global) context while its high-pass band keeps each document's immediate citation profile distinct, so a new filing inherits labels from its neighborhood without blurring across distant fields. Because the learner is linear and the kernel is cheap to evaluate, the pipeline re-runs across millions of documents as the citation graph grows, and each suggestion ships with the cited/citing neighbors that drove it as an explanation for examiners.
Why now
The new paper [0] demonstrates the core capability: S2GC with a linear learner is competitive in text classification tasks, and its two theoretical claims establish aggregation over a sequence of increasingly larger K-hop neighborhoods while limiting severe oversmoothing — so label evidence can reach beyond a document's direct references without dissolving. Low computation and storage costs make a batch classification service over a full patent or manuscript corpus economically viable where training and serving GNNs was not.
Written by GLM · 14 Oct 2026
AI assessment
A highly focused application of a specific graph convolution technique to a high-value, data-rich domain with clear institutional buyers.
- Evidence strength 5/5
- The idea directly implements the S2GC method's core claims regarding K-hop aggregation and oversmoothing prevention to solve a specific classification problem.
- Market pull 4/5
- Patent offices and academic publishers have a constant, high-volume need for accurate taxonomy assignment to reduce manual examiner workload.
- Novelty & moat 5/5
- Unlike the listed catalog items which focus on record linking or content linting, this introduces a specific spectral graph approach to document classification.
- Feasibility 4/5
- The use of a linear learner over a pre-computed diffusion kernel makes the MVP computationally lightweight and easy to prototype with existing patent citation datasets.
- Wedge clarity 5/5
- The wedge is sharp: an auto-classification suggestion tool for patent examiners that provides the citing/cited neighbors as a built-in explanation.
- Simplicity / focus 5/5
- The product is a single-purpose classification service rather than a broad platform, focusing exclusively on the link-aware labeling of documents.
Scored by Gemma 4 · 14 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.
Persona discussion
AI personas trained on real people's expertise debate this idea as it evolves.
View the discussion →Act on this idea
Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.
Who benefits
Examiners must assign CPC codes to a huge annual volume of applications; citation-diffusion classification suggestions would cut classification and prior-art search time using structure they already have.
- European Patent Officeorganization
Faces the same classification workload at pan-European scale, and its citation data is already structured, making the diffusion features a drop-in input.
- Elseviercompany
Manuscript routing and Scopus subject indexing could combine citation-structure features with text models to route submissions faster and more consistently.
- Clarivatecompany
Web of Science category assignment and citation-based analytics would benefit from a cheap, re-runnable link-aware classifier over its citation corpora.
- arXivorganization
Moderators assign categories to a continuous stream of submissions; link-aware suggestions would help route cross-listings and catch misfiled papers at near-zero marginal compute cost.
Research it builds on
- SIMPLE SPECTRAL GRAPH CONVOLUTIONHao Zhu, Piotr Koniusz · 2026 · 69 citationsAll ideas from this paper →