Source-linked AI summary
From Data Footprints to Researcher Fingerprints: Temporal Ego-centric Research Group Detection in Scientific Knowledge Graphs
Simone Angioni, Miriam Baglioni, Andrea Mannocci, Paolo Manghi
TL;DR
Understanding how researchers’ groups form and change over time matters for interpreting multidimensional researcher fingerprints, while peer review is difficult to sustain at scale. The paper applies an ego-centred Leiden-based method to Scientific Knowledge Graph data; in a 10-researcher evaluation, it achieved approximately 7.8/10 average accuracy and a median score of 8/10.
Problem
Research fingerprints require collaboration groups to be understood over time, while qualitative peer review is resource-intensive and difficult to sustain at scale.
Source: §1 · pp. 1–2Method
The method applies Leiden-based group detection to an individual researcher’s ego network and requires the ego researcher to appear in each identified group.
Source: §1 · p. 2Results
Approximately 7.8/10 average accuracy and a median score of 8/10 in a user evaluation involving 10 researchers.
Source: §6 · p. 9Takeaways & Limitations
Identified groups provide contextual structure for interpreting other fingerprint dimensions, rather than directly measuring researcher impact or merit or replacing peer review.
Source: §7 · p. 10Scalability to larger researcher populations and generalisability across disciplines remain to be assessed through broader user evaluation.
Source: §7 · p. 10
Abstract
from arXiv · showhide
Understanding who researchers collaborate with, when and how their research groups form, and how these evolve over time is a foundational prerequisite for constructing multidimensional, evidence-based researcher fingerprints -- a requirement that has become increasingly urgent in the context of ongoing reforms in research evaluation advocated by initiatives such as DORA, the Leiden Manifesto, and CoARA. In this paper, we propose an egocentric approach for fine-grained research group detection, operating at the scale of individual researchers' ego networks, as extracted from Scientific Knowledge Graphs (SKGs), and requiring the ego node to be present in all identified research groups -- a constraint that is not merely technical but reflects a commitment to researcher-centred analysis. Our approach builds on the Leiden algorithm as its computational core and is grounded in SKGs, which provide the heterogeneous, open scholarly metadata necessary to trace collaboration patterns at scale in a reproducible and transparent manner. We evaluate the proposed method through a user evaluation involving 10 researchers with diverse backgrounds and career stages, demonstrating its ability to surface meaningful and interpretable collaboration structures. The research groups identified constitute the structural scaffold upon which richer characterisations of scholarly identity -- including roles, seniority, contribution types, and publication patterns -- can be built in subsequent work, contributing to the broader goal of fairer and more transparent research assessment.
1. Introduction
Research evaluation reforms call for broader, evidence-based insights, while understanding a researcher’s collaboration history provides context for a multidimensional research fingerprint. This paper uses Scientific Knowledge Graphs and an ego-centred Leiden-based approach to identify research groups around individual researchers.
1. Introduction
- Peer review remains a gold standard but is resource-intensive and difficult to sustain at scale, motivating complementary data-driven insights into researchers’ activity.The paper situates this need amid evaluation reforms that look beyond publication-centred metrics toward broader contributions.Source: §1 · p. 1
- Understanding how research groups form and evolve provides the structural context needed to interpret researchers’ roles, outputs, impact, and other fingerprint dimensions.The paper presents group identification as a foundational layer for richer characterisations of scholarly identity.Source: §1 · p. 2
- Scientific Knowledge Graphs integrate heterogeneous scholarly metadata and relations, supporting large-scale, reproducible tracing of collaboration patterns.The paper identifies these open infrastructures as a basis for evidence-grounded researcher fingerprints.Source: §1 · p. 2
- Existing community-detection methods often partition global co-authorship networks, while the ego-membership constraint for scientific collaboration subgroups remains unexplored.The proposed researcher-centred constraint is intended to make identified groups meaningful to an individual’s collaborative experience and evolving role.Source: §1 · p. 2
- The proposed approach detects collaboration subgroups in an individual researcher’s ego network, requiring the researcher to belong to every identified group.It uses the Leiden algorithm as its computational core and is evaluated with 10 researchers.Source: §1 · p. 2
2. Related Works
Prior work spans community detection and research-team identification, but often uses static partitions or institution-, discipline-, or period-specific settings. This paper instead reconstructs a focal researcher’s evolving collaboration groups as contextual structures.
2. Related Works
- Community detection includes modularity, spectral, probabilistic, overlapping-community, and representation-learning approaches, with Leiden preferred over Louvain for partition refinement and connected communities.The literature also distinguishes disjoint from overlapping, static from dynamic, and topology-only from attributed or representation-learning settings.Source: §2 · p. 3
- Prior research-group studies use co-authorship structure, citation analysis, or project and contribution information, often within institutional, disciplinary, or formal team settings.These studies include methods for identifying groups from co-author choices, self-organising production units, and different university research-team types.Source: §2 · pp. 3–4
- Single partitions, local clusters, and period-specific teams are less suited to reconstructing a researcher’s career-spanning groups, especially for interdisciplinary, mobile, or senior scholars.Such histories may span multiple, partially overlapping groups, institutions, and periods.Source: §2 · p. 4
- The approach is ego-centred, treats communities as an intermediate structure, and tracks yearly group membership to identify multiple temporally persistent groups.This contrasts with a single static cluster or one-time graph partition.Source: §2 · p. 4
3. Data
The study uses the OpenAIRE Graph as its empirical foundation, selecting it for openness, coverage of non-traditional research outputs, and transparent curation. Research products are retrieved for each selected researcher using their ORCID identifier.
3. Data
- The OpenAIRE Graph integrates publications, data, software, and other research products through semantic links such as citations, affiliations, project participation, and funding.It aggregates and enriches metadata from over 2,100 direct sources and approximately 155,000 sources when indirect sources are included.Source: §3 · p. 4
- Researchers were selected by ORCID identifier to retrieve associated products from the OpenAIRE Graph 10.6.0 snapshot from the end of 2025.The graph includes works verified on ORCID profiles and works inferred by OpenAIRE enrichment algorithms.Source: §3 · p. 4
- The Graph was chosen for its open licence, inclusion of non-traditional outputs as first-class research products, and transparent, community-governed curation.These properties are presented as supporting reproducible and auditable fingerprint construction.Source: §3 · p. 4
4. Methodology
The methodology constructs an ego-centered co-authorship network, recursively detects its communities with Leiden, and assembles relevant collaborators into temporally coherent research groups. It addresses both incomplete neighborhood connectivity and the risk that a single partition conceals distinct groups.
4. Methodology
- The method combines ego-network construction, Leiden community detection, and temporal group assembly to identify research groups around an individual researcher.The ego node is required in each identified group.Source: §4 · p. 4
4.1. Ego Network Construction
- The induced weighted ego network restores co-author ties within the ego’s neighborhood, including collaborations that do not involve the ego directly.A standard ego network can underestimate or omit these ties, distorting the neighborhood’s community structure.Source: §4.1 · p. 5
4.2. Community Detection
- Recursive Leiden runs remove the ego’s community members after each iteration, revealing additional groups that a single hard partition could conflate or suppress.The first iteration captures the most prominent ego-centered community; later iterations progressively reveal others.Source: §4.2 · p. 6
4.3. Research Group Identification
- Each year, a collaborator qualifies for group membership when their co-authored products exceed a predefined share of the ego’s yearly research output.The predefined threshold is 30%.Source: §4.3 · p. 7
- Yearly members are assigned to community-based groups, then linked across years when temporal continuity holds or their instances remain structurally connected.If no community-specific group meets the minimum size, a sufficiently large fallback group of all yearly members is considered.Source: §4.3 · p. 7
- The output is a chronologically ordered set of groups represented by their time spans and members, capturing temporally persistent, community-coherent structures.Groups are filtered by minimum size, and groups with identical membership are merged across their observed intervals.Source: §4.3 · pp. 7–8
5. The Researcher Group Viewer
The Researcher Group Viewer builds and displays collaboration networks centered on an ORCID-identified researcher, with rapid API-based and more extensive precomputed options. Its visualizations let users explore group structure, outputs, and change over time.
Network construction
- Rapid mode retrieves ORCID-linked products and verified-ORCID co-authors through OpenAIRE APIs, while a requested precomputed network can also include works with inferred ORCIDs.The rapid mode may omit products not yet confirmed in ORCID profiles.Source: §5 · p. 8
Network visualization
- Authors appear as nodes, weighted co-authorships as edges, and community membership organizes collaborators around the highlighted central researcher.This view presents the broader collaboration landscape alongside specific research groups.Source: §5 · p. 8
Temporal exploration
- Year-by-year views and a timeline show collaboration changes and each detected group’s start and end years.The interface also displays associated product counts and types, authors, and weaker collaborative frameworks.Source: §5 · p. 8
Research outputs and links
- Edges distinguish publications, datasets, software, and other outputs, showing collaboration intensity through weights and product types through categories.Users can follow edges to related products and OpenAIRE Explore records.Source: §5 · pp. 8–9
6. Evaluation and Discussion
In a survey of 10 researchers, the method received average accuracy of approximately 7.8/10 (median 8/10), while feedback exposed identity-coverage and collaboration-context limits. Evaluators also identified temporal, thematic, and group-structure cues for future refinement.
6. Evaluation and Discussion
- Graphs used evaluators’ ORCID-linked research products, and feedback covered addition, removal, split, and move corrections plus 1–10 accuracy ratings.The evaluation included only products where the evaluator appeared as an author and co-authors with ORCID identifiers available in the Graph.Source: §6 · p. 9
- Accuracy averaged approximately 7.8/10 (median 8/10) across 10 evaluators, with seven assigning 8 and individual scores spanning 5–9.The evaluators had an average of 19 years of research experience.Source: §6 · p. 9
- Of 39 missing researchers, 16 lacked ORCID identifiers and 6 were already recognized in other research groups.The authors note that their approach could not identify researchers without ORCID profiles.Source: §6 · p. 9
- Removals often involved sporadic collaborations, annual reports, large shared datasets, or many-author publications, which may overstate stable collaboration.The authors retained these product types for computer-science-related researchers, while calling for cross-field analysis of which works indicate research groups.Source: §6 · pp. 9–10
- Two evaluators suggested splitting or merging groups, indicating occasional mismatches between detected boundaries and experts’ perceived collaboration contexts.The feedback included both failure to distinguish closely related communities and separation of groups experts viewed as belonging together.Source: §6 · p. 10
- Five evaluators suggested moving researchers whose group assignment did not match their interpretation of its thematic or temporal context.The researchers themselves were correctly identified in these cases.Source: §6 · p. 10
- Evaluators recommended characterizing groups by temporal continuity, topics, publication context, bridging researchers, stable cores, leaders, and formal ties.Other suggestions included visualizing overlapping memberships across years, enabling direct feedback, identifying key connectors, and letting users adjust group-boundary criteria.Source: §6 · p. 10
7. Conclusion and future works
The method supplies contextual structure for interpreting researcher fingerprints, not a measure of impact or merit or a substitute for peer review. The conclusion identifies data and disciplinary limits while outlining extensions to group characterization and career trajectories.
7. Conclusion and future works
- Detected groups provide context for interpreting research outputs and collaboration behaviour, rather than measuring researcher impact or merit or replacing qualitative peer review.The same output or behaviour may carry different significance across established groups, transitions, or newly formed interdisciplinary teams.Source: §7 · p. 10
- The authors plan to test scalability across larger researcher populations and generalisability across disciplines through a larger, more diverse user evaluation.Source: §7 · p. 10
- Dependence on available, curated ORCID identifiers can leave reconstructed networks partial because co-authors without verified identifiers are excluded.Source: §7 · p. 10
- Co-authorship alone can create spurious links from shared datasets, deliverables, or massively coauthored publications, motivating discipline-sensitive filtering or down-weighting.Filtering should avoid discarding relevant non-publication research products, including in law and the humanities.Source: §7 · pp. 10–11
- Future group descriptions may add thematic descriptors from topics, keywords, and Fields of Science to distinguish stable, topic-focused communities from heterogeneous or occasional collaborations.Stable cores, bridging authors, and institutional or formal collaboration contexts are also proposed as descriptors.Source: §7 · p. 11
- Evaluation in computer-science-related fields limits generalisability; algorithm behaviour in life sciences and social sciences remains to be assessed.Collaboration norms, group sizes, authorship conventions, and output types vary across disciplines.Source: §7 · p. 11
- Future work will examine leadership, contributory, and peripheral roles using authorship position, temporal consistency, and subgroup network centrality.Group longevity, membership stability, and authorship patterns are proposed as proxies for seniority and scientific maturity over time.Source: §7 · p. 11
- Future analyses will relate group output types to researcher role and seniority, and examine publication behaviour across collaborative contexts and as groups evolve.Planned output distinctions include publications, datasets, software, and other research products; publication patterns include venue choice, open access, and data sharing.Source: §7 · p. 11