Citation & Family-Grounded Clusters

• Citation and Family-Grounded Clusters are formal, data-driven methods that use direct citation, co-citation, and RPYS to map the intellectual structure of scientific literature.
• They employ algorithms such as spectral clustering, agglomerative merging, and hierarchical methods to achieve reproducible, quantifiable field segmentation.
• These methodologies reveal disciplinary genealogies by linking articles through citation networks and identifying enduring concept symbols across time.

Citation and Family-Grounded Clusters represent a set of formal, data-driven methodologies for mapping the intellectual organization of scientific literature using citation networks. These approaches leverage direct citation, co-citation, and extended spectroscopic techniques (e.g., RPYS) to reconstruct disciplinary “families,” reveal conceptual clusters, and uncover the evolution of scientific subfields at fine granularity. When instantiated as family-grounded clusters, citation-based methodologies parallel the logic of scientific descent, offering algorithmic analogues to traditional historiographies but based solely on article-level citation data—enabling reproducible, quantifiable classification of fields and subfields (Devarakonda et al., 2019, Comins et al., 2016, Gualdi et al., 2011).

1. Foundational Concepts and Motivations

Citation-based clustering methodologies draw on the hypothesis that the structure of citations encodes intellectual affinities and research descendancies. Direct-citation clusters (“family-grounded” clusters) comprise works that cite each other or are linked through chains of citations, often reflecting established communities of practice or research lineages. Co-citation clusters, on the other hand, are defined by papers that are jointly cited by later works and thus capture emergent conceptual linkages or latent thematic kernels that are not evident from direct citation alone (Devarakonda et al., 2019).

Reference Publication Years Spectroscopy (RPYS) introduces a third axis: by analyzing the temporal distribution of cited references within a domain, RPYS identifies historic “concept-symbols”—publications that function as anchoring intellectual sources across decades, differentiating them from short-lived research front peaks (Comins et al., 2016).

2. Formal Graph Representations and Clustering Algorithms

Three principal formalizations structure the implementation of citation/family-based clusters:

• Direct Citation Graph ($G_d = (V, E_d)$): Nodes $V$ are publications; edges $(i \to j)$ exist if $i$ cites $j$; adjacency matrix $A_{ij} = I(i \to j)$. For clustering, the undirected version $G_u$ includes an edge between $i$ and $j$ if at least one direction is present.
• Co-citation Graph ($G_c = (V', E_c, w)$): $V'$ contains highly cited papers. For each pair $(i,j)$, the co-citation weight is $w_{ij}^{co} = \sum_{k \in V} I(k \to i) I(k \to j)$, normalized by Salton’s cosine: $w_{ij} = w_{ij}^{co} / \sqrt{C_i C_j}$.
• Citation Backbone (Family Tree): Each paper is linked to a single “most impacting” parent. The core parent is determined by maximizing an impact score $I_{\alpha \to i}$, derived from weighted two-step random walks over common references (“author’s view”) and common citers (“reader’s view”); parameter $f$ typically set to $0.5$ (Gualdi et al., 2011).

Summary of clustering procedures:

Approach	Graph/Feature	Clustering Algorithm
Direct citation (family)	$G_u$ (undirected)	Spectral (e.g. Graclus)
Co-citation	$G_c$ (weighted)	Variable-level + agglomeration
Citation backbone	Tree (forest)	Time/branch-cut
RPYS-based	Spectrograms	Ward’s hierarchical

3. Implementation Protocols and Parameterization

Direct-Citation (“Family-Grounded”) Clustering

• Applied to an undirected citation graph, spectral methods such as Graclus are used to partition the literature. Normalized cut objectives and conductance (φ) evaluate cluster separation ($$\phi(S) = |\partial(S)|/\min(\text{vol}(S),2m-\text{vol}(S))$$).
• In large-scale computer science mapping, 20 clusters represented 2,685,356 publications, with cluster quality controlled by size balance (largest ≤10× smallest), stability, and conductance (median φ = 0.15 for clusters 0–18) (Devarakonda et al., 2019).

Co-Citation Clustering

• Highly cited nodes ($V'$: top 10th percentile) undergo weighted edge thresholding ($t$ increasing from 0.5 to 0.999 in steps), followed by connected-component extraction. Clusters of 100–200 nodes are retained, with larger components successively thresholded.
• Resulting subclusters are agglomeratively merged using maximum inter-cluster edge weight $W(X,Y)$ until a target number (e.g., 20) top-level clusters is achieved (Devarakonda et al., 2019).

Citation Backbone and Family-Tree Clustering

• Each paper with parents is connected to its most impacting parent, as quantified mutually by two-step similarities over references and citers. The backbone is a directed forest, yielding a tree-like family structure (Gualdi et al., 2011).
• Clusters (fields/subfields) are extracted by either time cut (all nodes after year $Y_c$) or branch-depth cut (splitting subtrees at branching points of depth $S$), with exclusivity $E$ and topical coherence ($N_P$) used to tune parameters for quality.

RPYS and Multi-RPYS

• For a document set $P$, counts $C(t)$ of cited references by year $t$ are analyzed, transformed via a five-year median window into deviation series $D(t)$ and rank spectrograms $S(t)$. Multi-RPYS partitions $P$ into $m$ intervals for comparative historicity, flagging “persistent peaks” as concept-symbols if widely distributed and long-lived (Comins et al., 2016).
• Feature vectors $v_J$ (RPYS ranks by year for journal $J$) provide similarity/distance metrics (e.g., squared Euclidean), feeding hierarchical clustering (Ward’s criterion) to form intellectual families.

4. Field Mapping, Disciplinary Structure, and Cluster Evaluation

Quantitative and qualitative evaluation employs several complementary metrics:

• Conductance (φ): Lower values indicate better internal cohesion for direct-citation clusters.
• Exclusivity (E): Compares intra-cluster citation localization to random baseline.
• Effective PACS number ($N_P$): Topical homogeneity; lower $N_P$ denotes more coherent subject focus.
• Fractal analysis: Box-covering of citation backbones reveals self-similar, hierarchical development.
• Cluster-ASJC mapping: Cross-linkage of article-level clusters to Scopus All Science Journal Classification (ASJC) major/minor categories enables comparison with journal-based taxonomies (Devarakonda et al., 2019).

Cluster examples illustrate both cross-category (interdisciplinary) and inward-looking (core field) structures. For instance, a direct-citation cluster may unify biochemistry, computational mathematics, genetics, and statistics, reflecting a computational biology/math nexus, while a co-citation cluster nucleated by landmark robotics papers may traverse multiple direct-citation clusters.

5. Empirical Studies and Case Findings

Case studies validate family-grounded clustering in multiple domains:

• Computer Science (DBLP+Scopus): Direct-citation clusters align with broad fields (hardware, software, theory), bridge to external disciplines (biology, neuroscience), and correspond well to ASJC minor category distributions. Co-citation clusters serve as thematically sharper “hotspots,” often nested within direct-citation blocks (Devarakonda et al., 2019).
• Physics (APS Network): The citation backbone yields trees of subfields; cluster quality (in $E$ and $N_P$) is superior to random or naive long-path controls. The backbone exhibits robust fractal dimensions, supporting the family-descent analogy (Gualdi et al., 2011).
• Information Science (RPYS): Multi-RPYS segmentation clusters journals into coherent intellectual lineages: e.g., “Librarianship & Library Systems” versus “Theoretical & Quantitative Information Science.” Persistent concept-symbol years distinguish long-term field-defining citations from transient research front phenomena (Comins et al., 2016).

Domain	Approach	Key Observations
Computer Science	Direct, co-citation	Direct: broad, cross-labeled; co-citation: focused, conceptual
Physics	Citation backbone	Higher exclusivity, coherence, and self-similarity
Information Sci	RPYS, multi-RPYS	Long-term concept-symbols cluster journals with historical ties

6. Comparative Insights and Complementarity

Direct-citation (family) clusters are generally broader, mapping communities rooted in explicit scholarly descent, while co-citation clusters isolate emergent or niche conceptual domains. Co-citation may capture developments that retroactively bind papers not directly interlinked. When integrated, both modes achieve a dual mapping—direct-citation reveals established "family trees," and co-citation surfaces the intellectual crosswinds and field innovations. RPYS and its multi-interval extension provide orthogonal clustering axes, focusing on temporally persistent intellectual symbols and enabling transversal clustering by shared historic foundations (Devarakonda et al., 2019, Comins et al., 2016, Gualdi et al., 2011).

7. Significance, Limitations, and Interpretive Considerations

Citation and family-grounded clusters provide high-resolution, reproducible frameworks for intellectual field mapping, disciplinary genealogy, and algorithmic historiography. They enable rigorous field subdivision, capture interdisciplinary bridges, and can recover latent conceptual structures overlooked by journal-level or keyword-based approaches.

A plausible implication is that, while citation-based methods operationalize dependency and intellectual proximity, they do not necessarily correspond to epistemic “truth” or objective field boundaries. Noise from citation conventions, data artifacts, or extrinsic factors (e.g., journal policies) may affect cluster topology. Direct-citation approaches particularly encode communities of practice, while co-citation and RPYS methodologies highlight emergent associations and long-term conceptual roots, respectively.

These methods, jointly applied, delineate the evolving topology of science, supporting both quantitative evaluation and the historiographic reconstruction of scientific development at scale (Devarakonda et al., 2019, Comins et al., 2016, Gualdi et al., 2011).