Unveiling the Hidden Dynamics of Knowledge Graphs: The Role of Superficiality in Structuring Information

ORCID_LOGO based on reviews by Mateusz Wilinski, Tamao Maeda and Abiola Akinnubi
A recommendation of:

The Structure and Dynamics of Knowledge Graphs, with Superficiality

Data used for results
Codes used in this study
Scripts used to obtain or analyze results


Submission: posted 16 May 2023, validated 18 May 2023
Recommendation: posted 01 June 2024, validated 05 June 2024
Cite this recommendation as:
Sueur, C. (2024) Unveiling the Hidden Dynamics of Knowledge Graphs: The Role of Superficiality in Structuring Information. Peer Community in Network Science, 100113. 10.24072/pci.networksci.100113


Knowledge graphs [1–4] represent structured knowledge using nodes and edges, where nodes signify entities and edges denote relationships between these entities. These graphs have become essential in various fields such as cultural heritage [5], life sciences [6], and encyclopedic knowledge bases, thanks to projects like Yago [7], DBpedia [8], and Wikidata [9]. These knowledge graphs have enabled significant advancements in data integration and semantic understanding, leading to more informed scientific hypotheses and enhanced data exploration.

Despite their importance, understanding the topology and dynamics of knowledge graphs remains a challenge due to their complex and often chaotic nature. Current models, like the preferential attachment mechanism, are limited to simpler networks and fail to capture the intricate interplay of diverse relationships in knowledge graphs. There is a pressing need for models that can accurately represent the structure and dynamics of knowledge graphs, allowing for better understanding, prediction, and utilisation of the knowledge contained within them.

The paper by Lhote, Markhoff, and Soulet [10] introduces a novel approach to modelling the structure and dynamics of knowledge graphs through the concept of superficiality. This model aims to control the overlap between relationships, providing a mechanism to balance the distribution of knowledge and reduce the proportion of misdescribed entities. This is the first model tailored specifically to knowledge graphs, addressing the unique challenges posed by their complexity and diverse relationship types. The innovation lies in the introduction of superficiality, a parameter that governs the probability of adding new entities versus enriching existing ones within the graph. This model not only addresses the multimodal probability distributions observed in real KGs but also offers a more granular understanding of the knowledge distribution, particularly the presence of misdescribed entities. The authors validated their model against three major knowledge graphs: BnF, ChEMBL, and Wikidata. The results demonstrated that the generative model accurately reproduces the observed distributions of incoming and outgoing degrees in these knowledge graphs. The model successfully captures the multimodal nature and the irregularities in the degree distributions, especially for entities with low connectivity, which are typically the majority in a knowledge graphs.

One significant finding is the impact of superficiality on the level of misdescribed entities. The study revealed that lower superficiality leads to a more uniform distribution of relationships across entities, thus reducing the number of entities described by few relationships. Conversely, higher superficiality results in a higher proportion of entities with minimal descriptive facts, reflecting a paradox where increasing the volume of knowledge does not necessarily reduce the level of ignorance. The authors also conducted an ablation study comparing their model to traditional models like Barabási-Albert [11] and Bollobás [12]. The results showed that the proposed multiplex model with superficiality parameters consistently outperformed these traditional models in accurately reflecting the characteristics of real-world knowledge graphs. 

This research provides a groundbreaking approach to understanding and modelling the structure and dynamics of knowledge graphs. By introducing superficiality, the authors offer a new lens through which to examine the distribution and organisation of knowledge within these complex structures. The model not only enhances our theoretical understanding of knowledge graphs but also has practical implications for improving data storage, query optimisation, and the robustness of knowledge induction processes.

The introduction of superficiality opens several avenues for future research and application. One potential direction is refining the model to account for localised perturbations in smaller knowledge graphs or specific domains within larger knowledge graphs. Additionally, longitudinal studies could further elucidate the evolution of superficiality over time and its impact on the quality of knowledge representation. Another promising area is the application of this model in real-time knowledge graphs management systems. By adjusting superficiality parameters dynamically, it may be possible to optimise the balance between entity enrichment and the introduction of new entities, leading to more robust and accurate knowledge graphs. In the broader context of knowledge engineering and data science, this model offers a framework for exploring the vulnerability of knowledge graphs and their susceptibility to various types of biases and inaccuracies. This understanding could lead to the development of more resilient knowledge systems capable of adapting to new information while maintaining a high level of accuracy and coherence.

Overall, the concept of superficiality and the associated generative model represent significant advancements in the study and application of knowledge graphs, promising to enhance both our theoretical understanding and practical capabilities in managing and utilising these complex data structures. It would be interesting to see how this can be extended to domains in social network analyses [13,14].


​1.     Nickel M, Murphy K, Tresp V, Gabrilovich E. 2015 A review of relational machine learning for knowledge graphs. Proceedings of the IEEE 104, 11-33.

2.     Ehrlinger L, Wöß W. 2016 Towards a definition of knowledge graphs. SEMANTiCS (Posters, Demos, SuCCESS) 48, 2.

3.     Hogan A et al. 2021 Knowledge graphs. ACM Computing Surveys (Csur) 54, 1-37.

4.     Ji S, Pan S, Cambria E, Marttinen P, Philip SY. 2021 A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems 33, 494-514.

5.     Bikakis A, Hyvönen E, Jean S, Markhoff B, Mosca A. 2021 Special issue on semantic web for cultural heritage. Semantic Web 12, 163-167.

6.     Santos A et al. 2022 A knowledge graph to interpret clinical proteomics data. Nature biotechnology 40, 692-702.

7.     Suchanek FM, Kasneci G, Weikum G. 2007 Yago: a core of semantic knowledge. pp. 697-706.

8.     Auer S, Bizer C, Kobilarov G, Lehmann J, Cyganiak R, Ives Z. 2007 Dbpedia: A nucleus for a web of open data. pp. 722-735. Springer.

9.     Mora-Cantallops M, Sánchez-Alonso S, García-Barriocanal E. 2019 A systematic literature review on Wikidata. Data Technologies and Applications 53, 250-268.

10.  Lhote L, Markhoff B, Soulet A. 2023 The Structure and Dynamics of Knowledge Graphs, with Superficiality. arXiv, ver. 3 peer-reviewed and recommended by Peer Community in Network Science.

11.  Barabási A-L, Albert R. 1999 Emergence of scaling in random networks. science 286, 509-512.

12.  Bollobás B, Borgs C, Chayes JT, Riordan O. 2003 Directed scale-free graphs. pp. 132-139. Baltimore, MD, United States.

13.  Sueur C, King AJ, Pelé M, Petit O. 2013 Fast and accurate decisions as a result of scale-free network properties in two primate species. In Proceedings of the European conference on complex systems 2012 (eds T Gilbert, M Kirkilionis, G Nicolis), pp. 579-584.

14.  Romano V, Shen M, Pansanel J, MacIntosh AJJ, Sueur C. 2018 Social transmission in networks: global efficiency peaks with intermediate levels of modularity. Behav Ecol Sociobiol 72, 154.

Conflict of interest:
The recommender in charge of the evaluation of the article and the reviewers declared that they have no conflict of interest (as defined in the code of conduct of PCI) with the authors or with the content of the article. The authors declared that they comply with the PCI rule of having no financial conflicts of interest in relation to the content of the article.
The authors declare that they have received no specific funding for this study

Evaluation round #1

DOI or URL of the preprint:

Version of the preprint: 2

Author's Reply, 31 May 2024

Decision by ORCID_LOGO, posted 31 Jul 2023, validated 02 Aug 2023

Dear authors,

Pleasé find three reviews that can help you to enhance your preprint. We hope to see a revision soon.

Thanks for submitting your preprint in PCI.


Cédric Sueur 

Reviewed by , 03 Jul 2023

Reviewed by , 30 Jul 2023

Overall the work is interesting and potentially useful in improving data storage or to evaluate the robustness of existing knoewledge graphs. Couple of questions and remarks on my side:

1. Authors could spend a bit more time when describing the whole concept of knowledge graphs. Although I do not know what types of journals Authors are aiming, some extra figures and examples could make the whole manusciprt much clearer for a non-specialist or a general reader.

2. What model do Authors use when they say Bollobas? Bollobas-Riordan? Though there is one citation, it is not referred to in the description of the experiment (comparing different models).

3. I am slightly puzzled by the results on comparing Authors' model with BA (their fits to data). Isn't BA a special case of the new described model? How can it be better for any instance then (outgoing ChEMBL for example)? Does it even make sense to make a comparison in this case? Is it similar for Bollobas?

4. Could Authors add the other models' fits to Fig. 1?

5. Authors claim, unless I missunderstood something, that their model capture the multimodality, but Fig. 1 does not seem to support that, especially in all the outgoing cases. There seem to be some other effect at play. That is why seeing how does the shape compare to other models would be useful (see previous point).

Minor comments:

1. Multiplex is not the same as multilayer, it is a special case of the latter.

2. Shouldn't the comparison between different models go to the main text?

3. The references use different styles: sometimes we have "et al." after the first author, sometimes after four authors and sometimes we have all five authors mentioned. I would suggest to "unify" it.

Reviewed by , 29 Jul 2023

User comments

No user comments yet