Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 78 of the 78 references it cites were found in a registry or an index; none of the 58 sources its claims cite is checked for support yet.

A Critical Review on Citation Recommendation Systems Through a Graph-Centric Approach

Open access
Kefalas, Pavlos, Ali, Zafar, Manolopoulos, Yannis · 2026 · SN Computer Science

Abstract

The rapid growth of digital scholarly content has made citation recommendation central to literature discovery. This survey organizes the field through a graph-centric lens, unifying unipartite, bipartite, and k -partite formulations and the typed relations that connect papers, authors, venues, concepts, and time. Despite growing research, several critical gaps remain in citation recommendation: (i) heterogeneous relations are underutilized, limiting the ability to model complex interactions; (ii) temporal dynamics and real-time updates are rarely incorporated, restricting adaptivity; (iii) social and serendipity signals (e.g., co-reading, communities, novelty/diversity) are largely neglected, reducing user-centric utility; (iv) cross-lingual and cross-domain transfer is underexplored, limiting global applicability; and (v) evaluation practices remain fragmented, emphasizing accuracy over transparency, diversity, and robustness. We consolidate datasets, tasks, and metrics, and outline a roadmap emphasizing relation-weighted heterogeneous graphs, temporally adaptive learners, cross-lingual alignment, explanation-aware objectives, and deployable pipelines with open artifacts. This graph-centric perspective surfaces actionable gaps and opportunities for building citation recommenders that are accurate, transparent, and production-ready.

Overview

CoChat Summary

TL;DR

This survey reframes citation recommendation through a graph-centric lens, synthesizing uni-, bi-, and k-partite formulations, typed relations, and temporal dynamics across 56 included studies. It identifies five key gaps—relation-aware modeling, temporal/streaming updates, social/serendipity signals, cross-lingual/cross-domain generalization, and fragmented evaluation—and offers a roadmap toward reproducible, deployable systems.

Key findings

  • The survey’s synthesis highlights five key findings: heterogeneous relations benefit from relation-aware weighting; temporal evolution and streaming updates are underutilized; social and serendipity signals are often neglected; cross-lingual and cross-domain generalization remains limited; and evaluation practices are fragmented, focusing mainly on accuracy over diversity, fairness, robustness, and deployability.
  • The PRISMA-guided search retrieved 236 records; after removing 78 duplicates, 158 unique studies were screened, 102 were excluded at title/abstract, and 56 underwent full-text assessment and were included.
  • Data sparsity remains severe: systems routinely face interaction matrices/tensors with >99% missing entries; for a 3D tensor with n=1000 users, m=2000 papers, k=300 time periods, and r=1000 interactions, density ≈ 0.000167% and sparsity ≈ 99.999833%.
  • Reproducibility and deployment gaps persist: only 5 out of 47 studies made their code publicly accessible, and real-world platforms such as Mendeley, PubMed, and ResearchGate continue to rely on traditional information filtering rather than state-of-the-art citation recommendation methods.

Methods

  • PRISMA-guided literature survey searching DBLP, IEEE Xplore, ACM Digital Library, Scopus, Web of Science, arXiv, and OpenAlex for works published between January 2010 and October 2025, with queries targeting citation recommendation, graph/GNN/HIN/meta-path, transformer/LLM/SciBERT/SPECTER, RAG, and cross-lingual/multilingual alignment.
  • After deduplication, 158 unique records were screened by title/abstract, followed by full-text assessment; two reviewers applied inclusion/exclusion criteria, resolving disagreements by discussion, yielding 56 included studies.
  • For each included paper, the authors extracted task type (global vs. local), input modalities, graph construction, modeling approach, datasets, and evaluation; the analysis uses a graph-centric taxonomy unifying uni/bi/k-partite formulations and typed edges, plus a diagnostic synthesis of five problem axes. The paper states it is a survey and does not present new experimental results or case studies.

Limitations

  • The paper states it is a survey and does not present new experimental results or case studies; it also states that it does not provide empirical benchmarks.
  • The paper states that evaluation and reproducibility are fragmented in the field, with accuracy favored over transparency, diversity, and robustness, and that only 5 out of 47 studies made code publicly accessible.
  • Evident from the text: the review is restricted to works published between January 2010 and October 2025 and excludes non-English/abstract-only items and generic paper recommendation without a citation objective, so its synthesis depends on prior literature and reported metrics.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s42979-026-04848-2

Cites

78 of 78 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Annual publication trends highlighting the growth of the citation recommendation domain (yellow) relative to overall computer science research (cyan), illustrating the increasing research focus in this area
Fig. 1 Annual publication trends highlighting the growth of the citation recommendation domain (yellow) relative to overall computer science research (cyan), illustrating the increasing research focus in this areaCC-BY · From the PDF
Image 2
CC-BY · From the PDF
Image 3
CC-BY · From the PDF
Fig. 4 Representation of k -partite citation graph relationships
Fig. 4 Representation of k -partite citation graph relationshipsCC-BY · From the PDF
Image 5
CC-BY · From the PDF
Fig. 6 Depiction of repetitive user preference dynamics that limit serendipity, showing how overly personalized recommendations can reinforce narrow consumption patterns
Fig. 6 Depiction of repetitive user preference dynamics that limit serendipity, showing how overly personalized recommendations can reinforce narrow consumption patternsCC-BY · From the PDF
Fig. 7 Bipartite graph illustrating how user interactions and social network connections influence recommendations, highlighting the role of social signals in personalization
Fig. 7 Bipartite graph illustrating how user interactions and social network connections influence recommendations, highlighting the role of social signals in personalizationCC-BY · From the PDF
Fig. 8 Example of a bipartite recommendation system providing both personalization and justifica -tion, showing top recommendations for a target user along with the rationale derived from connected users
Fig. 8 Example of a bipartite recommendation system providing both personalization and justifica -tion, showing top recommendations for a target user along with the rationale derived from connected usersCC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 8 Not valid · 38 Not checked · 68 Own results and interpretations

15 more claims

Create a free account to see all 114 claims and to verify the 55 references still open.

See all claims — free account
  • Probably not a substantial claim. 82
  • It may state two separate claims as one. 45

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 10 · the rest aren’t loaded

5 more findings

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Show all 78 references

Research works printed but not linked: 5

  1. Buckley C, Voorhees EM. Evaluation of retrieval systems. In Proceedings of the 27th ACM annual international conference on research & development in information retrieval (SIGIR). 2004;1–12.
  2. Hidasi B, Karatzoglou A, Baltrunas L, Tikk D. Session-based recommendations with recurrent neural networks. In Proceedings of the 4th international conference on learning representations (ICLR). 2016.
  3. Liu T-Y. Learning to rank for information retrieval. Berlin: Springer; 2009.
  4. Rohatgi S. Design and data mining techniques for large-scale scholarly digital libraries and search engines. University Park: The Pennsylvania State University; 2023.
  5. Salton G, McGill MJ. Introduction to modern information retrieval. Columbus: McGraw-Hill; 1988.

Cited by

Metadata

Authors
Kefalas, P., Ali, Z., & Manolopoulos, Y.
Affiliations
  • Aristotle University of Thessaloniki (GR)
  • Southeast University (CN)
  • University of Nicosia (CY)
Corresponding author
Manolopoulos, Yannis
Venue
SN Computer Science
Published
2026-03-09
Volume
7
Issue
3
ISSN
2661-8907 (ISSN-L)
OpenAlex
W7134273351
Cited by
4
References
78
Citations by year
2026: 4
FWCI
10.52
Type
journal article
Language
en
Open access
Licence
Locations
SN Computer Science (publishedVersion, cc-by)
Funders
University of Nicosia
Topics
  • Physical Sciences › Computer Science › Information Systems › Recommender Systems and Techniques
  • Social Sciences › Library and Information Sciences › scientometrics and bibliometrics research
  • Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Complex Network Analysis Techniques
Keywords
citation recommendation, graph neural networks, heterogeneous graphs, bipartite graphs, k-partite graphs, temporal dynamics, shared reading, serendipity, cross-lingual transfer, cross-domain transfer, recommendation evaluation, literature discovery, scientific information retrieval

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.