Composite Cartel Index (CCI) Overview

• Composite Cartel Index (CCI) is a community-level anomaly score that measures unusual citation patterns in scholarly knowledge graphs.
• It combines multiple standardized signals—internal density, citation inflation, reciprocity, semantic superficiality, degree assortativity, and PageRank drop—to quantify anomalies.
• CCI functions as a ranked audit queue to prioritize communities exhibiting dense internal citations that may indicate suspicious structural-semantic behavior.

The Composite Cartel Index (CCI) is a community-level anomaly score for citation communities in a scholarly knowledge graph (SKG). It is designed to quantify how much a given group of papers resembles a citation cartel rather than a normal research area. In the formulation introduced in "Trust-Aware Citation Cartel Ranking in Scholarly Knowledge Graphs" (Gupta et al., 7 Jul 2026), a citation cartel is not treated as a proven case of misconduct, but as a suspicious structural–semantic pattern: a compact paper community with unusually heavy internal citation, elevated reciprocity, semantically shallow citation intent, locally assortative degree structure, and a marked loss of apparent influence when low-trust citations are downweighted. CCI therefore functions as a ranked audit queue for curator inspection rather than a verdict mechanism (Gupta et al., 7 Jul 2026).

1. Definition and rationale

CCI exists because citation-based systems often treat each citation as an equal signal of scholarly influence, even though citations can encode materially different relations, including direct method use, result comparison, broad background, or weak ceremonial acknowledgement. The motivating claim is that dense internal citation alone is not suspicious, because legitimate research communities are also modular and dense. Likewise, single-feature heuristics such as reciprocity alone or a high fraction of background citations alone are treated as inadequate. The proposed alternative is a multi-signal score that combines structural and semantic anomalies at the community level (Gupta et al., 7 Jul 2026).

The pipeline operates on a DBLP-derived citation graph with 500,000 papers and 4.87M citation edges. Communities are detected with Louvain, and each community is then characterized by six features: internal density, citation inflation, reciprocity, semantic superficiality, degree assortativity, and trust-weighted PageRank shift. CCI is computed by z-scoring these features across communities and averaging them. This ranking is explicitly intended to prioritize communities that warrant closer inspection, not to identify misconduct automatically (Gupta et al., 7 Jul 2026).

A central conceptual distinction is between ordinary dense specialization and structural–semantic abnormality. The framework assumes that suspiciousness emerges when multiple signals align: many more internal citations than expected under a degree-product baseline, unusually high mutual citation, a large fraction of Background or Perfunctory/Ceremonial citations, positive local assortativity in a globally disassortative graph, and a drop in PageRank when low-trust edges are downweighted. This suggests that CCI is best understood as a composite deviation score relative to the background citation ecology rather than as a classifier of intent.

2. Formal specification

Let $G=(V,E)$ be a directed citation graph, and let $C \subseteq V$ denote a Louvain community. For each community, the method computes six raw features $f_i(C)$, standardizes each feature across communities, and averages the resulting z-scores. The normalization is

$z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$

where $\mu_i$ and $\sigma_i$ are the mean and standard deviation of feature $i$ over the candidate communities. The Composite Cartel Index is then

$$\mathrm{CCI}(C) = \frac{1}{6}\sum_{i=1}^{6} z_i(C).$$

The six feature dimensions are:

1. Internal directed density
2. Citation inflation
3. Internal reciprocity
4. Semantic superficiality
5. Degree assortativity on log-degree
6. PageRank drop from unweighted to trust-weighted PageRank

This construction has no intrinsic bound. The paper states that, in practice, CCI is distributed roughly like a zero-centered Gaussian mixture across communities, and that it is used purely as a ranking rather than with a fixed threshold (Gupta et al., 7 Jul 2026).

The units of analysis are Louvain communities on the full citation graph. Communities are treated as isolated candidate communities, and the framework does not use overlapping clusters. The degree-based expectation for internal citation is based on a directed configuration-style null model, which is intended to control for community size and scale-free degree effects. This is significant because the methodology aims to isolate excess internal citation rather than density that is merely expected from node degree.

3. Structural and semantic components

The six CCI components combine graph topology, citation semantics, and trust-aware centrality. Their joint use is the defining feature of the index.

Component	Definition
Internal density	$m_C / (n_C(n_C-1))$
Citation inflation	$m_C / \mathbb{E}[m_C]$
Reciprocity	Fraction of internal edges in mutual pairs
Semantic superficiality	Fraction of typed internal edges labeled Background or Perfunctory/Ceremonial
Degree assortativity	Pearson correlation on source/target log-degrees over internal edges
PageRank drop	Mean difference between unweighted and trust-weighted PageRank

For a community $C \subseteq V$0, let $C \subseteq V$1, let $C \subseteq V$2, and let $C \subseteq V$3. The internal directed density is

$C \subseteq V$4

The citation inflation feature compares the observed number of internal citations with a degree-product expectation:

$C \subseteq V$5

with

$C \subseteq V$6

This null model is meant to distinguish excessive within-community citation from the citation concentration that would already be expected in a scale-free graph.

The internal reciprocity feature captures the fraction of internal edges that belong to a mutual pair. If

$C \subseteq V$7

then

$C \subseteq V$8

Because citation graphs are globally almost acyclic and low in reciprocity, local deviations are treated as meaningful anomaly signals.

The semantic superficiality feature uses typed internal edges only. If $C \subseteq V$9 is the set of internal edges with citation-intent labels, and if $f_i(C)$0 is the subset labeled Background or Perfunctory/Ceremonial, then

$f_i(C)$1

Untyped edges are ignored for this feature; they neither increase nor decrease the superficiality value directly.

The degree assortativity feature is computed on log-degree over internal edges. For each internal edge $f_i(C)$2,

$f_i(C)$3

and the feature is the Pearson correlation

$f_i(C)$4

The paper motivates this with the empirical observation that the DBLP citation graph is globally disassortative, so locally positive assortativity is anomalous.

Finally, the PageRank drop feature measures how much a community’s apparent influence depends on citations that receive low semantic trust weight. Let $f_i(C)$5 and $f_i(C)$6 denote the unweighted and trust-weighted PageRank values of node $f_i(C)$7. Then

$f_i(C)$8

A larger positive value indicates that the community loses more PageRank mass when low-trust citations are downweighted (Gupta et al., 7 Jul 2026).

4. Trust-aware citation semantics

The semantic layer of CCI is built from a teacher–student setup for citation intent classification. A GPT-4.1-mini teacher labels 205,897 citation pairs using six categories: Background, Method, Result/Comparison, Support, Contrast/Criticism, and Perfunctory/Ceremonial. A SciBERT student, specifically allenai/scibert_scivocab_uncased, is trained on these labels with a split of 164,717 training edges, 20,590 validation edges, and 20,590 test edges. Reported test metrics are accuracy 0.775, Macro-F1 0.574, Weighted-F1 0.775, Perfunctory/Ceremonial F1 0.473, and Contrast/Criticism F1 0.630. The trained model is then applied to 2,043,874 unique edges, corresponding to 42% of all edges, using targeted sampling that emphasizes random edges, edges in dense or high-inflation communities, reciprocal edges, low-similarity edges, and edges touching candidate communities (Gupta et al., 7 Jul 2026).

The paper defines a semantic trust weight $f_i(C)$9 for each citation edge. The mapping is:

Intent label	Weight
Method	1.0
Result/Comparison	0.7
Support	0.5
Contrast/Criticism	0.3
Background	0.2
Perfunctory/Ceremonial	0.1

Untyped edges receive weight 1.0 for PageRank so that missing labels do not automatically penalize papers. The trust-weighted PageRank is defined by

$z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$0

with damping factor $z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$1, stated as typically $z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$2. The corresponding unweighted PageRank uses unweighted out-degree in the transition probabilities.

This trust model is explicitly semantic, not probabilistic. A Method citation is treated as high-trust because it reflects direct methodological, tool, or dataset dependence; Background and Perfunctory/Ceremonial citations are treated as low-trust because they are broad, weak, generic, or ceremonial. In CCI, these labels influence community scoring in two distinct ways: directly through semantic superficiality, and indirectly through PageRank drop. A plausible implication is that the framework separates citation quantity from citation functional strength, which is precisely the distinction the paper argues is missing from citation-cartel detection based on topology alone.

5. Empirical behavior and validation

The highest-ranked community reported in the study contains 1,079 papers and 8,603 internal citations, with 254.3x more internal citations than expected and 64.2% of internal citations classified as superficial. Its reported CCI is 1.863. The other top-ranked communities demonstrate that high CCI can arise from different combinations of inflation and superficiality rather than from a single dominant signal (Gupta et al., 7 Jul 2026).

Rank	Community summary	CCI
1	Community 14: 1,079 papers, 8,603 internal edges, inflation 254.3, superficiality 0.642	1.863
2	Community 15: 22,125 papers, 138,436 internal edges, inflation 29.2, superficiality 0.800	0.441
3	Community 0: 16,575 papers, 68,851 internal edges, inflation 37.7, superficiality 0.705	0.402
4	Community 5: 6,960 papers, 66,597 internal edges, inflation 29.4, superficiality 0.539	0.373
5	Community 13: 12,052 papers, 122,138 internal edges, inflation 37.2, superficiality 0.850	0.234

The paper compares full CCI with several scoring variants: density only, inflation only, reciprocity only, structural-only CCI, semantic-only superficiality, and random. Reported correlations and top-$z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$3 overlaps with the full ranking show that no single feature reproduces the CCI ordering. The strongest variant is structural-only CCI, with Spearman 0.875, P@5 0.80, and P@10 0.70, but even this differs materially from the full index. Semantic-only superficiality reaches Spearman 0.680, P@5 0.40, and P@10 0.90. By contrast, reciprocity only yields Spearman 0.068, which reinforces the paper’s argument against single-heuristic detection (Gupta et al., 7 Jul 2026).

Leave-one-feature ablations likewise indicate that the ranking remains correlated when any single feature is removed, but that top-ranked membership changes, especially when semantic superficiality or PageRank drop is omitted. Reported Spearman correlations range from 0.875 to 0.998 depending on the removed feature, with noticeable changes in P@5 and P@10. This is used to support the claim that CCI depends on multiple partly independent signals rather than being reducible to density, inflation, or semantics alone.

The study also performs an edge excision stress test. Removing all 404,625 internal edges in the top five CCI communities leaves 94.1\% of nodes in the giant component, whereas matched random deletions of the same number of edges leave 99.95–99.96\% of nodes in the giant component. The reported interpretation is that these high-CCI communities form locally cohesive but globally peripheral structures. This suggests that the framework is identifying graph regions with unusually concentrated internal connectivity that do not behave like random dense subgraphs.

6. Interpretation, operational use, and limitations

CCI is intended for curator-facing ranking. To compute it on a new scholarly graph, the paper describes an implementation sequence: prepare a directed citation graph; run Louvain or similar modularity-based community detection; sample edges for intent annotation; train or obtain a citation-intent classifier; infer edge labels; assign trust weights; compute density, inflation, reciprocity, degree assortativity, and superficiality per community; compute unweighted and trust-weighted PageRank; derive PageRank drop; z-score all six features; and rank communities by the resulting average. The stated inputs and assumptions are a sufficiently large directed citation graph, citation-intent labels for at least a substantial subset of edges, and a reasonably accurate intent classifier, described as macro-F1 $z_i(C) = \frac{f_i(C)-\mu_i}{\sigma_i},$4. The reported system handles 500k papers and 4.87M edges, with intent inference identified as the main cost, limited to roughly 2M edges through targeted sampling (Gupta et al., 7 Jul 2026).

Several interpretive cautions are explicit. First, high CCI does not imply misconduct; it indicates a community whose structural and semantic profile is sufficiently unusual to merit audit. Second, CCI is sensitive to the quality of the citation-intent classifier. The paper notes that semantic features rely on a SciBERT model trained on LLM-generated labels and that macro-F1 is modest, although aggregate statistics over large communities are expected to be more stable than individual edge labels. Third, results depend on community detection sensitivity: using Louvain fixes the unit of analysis, and different methods or resolutions would change the candidate communities and possibly the ranking. Fourth, the pipeline is domain specific to the DBLP computer-science setting, and other fields may exhibit different citation norms and intent distributions (Gupta et al., 7 Jul 2026).

The paper also identifies several possible extensions: richer text context for intent modeling, temporal dynamics such as bursts of within-group citation over short windows, adaptation to other relational or trust graphs beyond citations, and alternative null models for expected internal citation. These proposals imply that CCI is best regarded as a modular framework rather than a final settled metric.

A recurrent misconception addressed by the framework is that a citation cartel can be detected from density alone. The study argues against this by showing that legitimate research areas can also be dense and modular, and by demonstrating empirically that density-only and inflation-only rankings diverge from full CCI. A second misconception is that semantic shallowness alone is sufficient. The reported results show that semantic-only superficiality overlaps with, but does not reproduce, the CCI ranking. The operational interpretation is therefore narrow but precise: CCI ranks communities that are simultaneously dense, inflated, reciprocal, semantically superficial, locally assortative, and trust-sensitive, thereby prioritizing suspicious citation structures for human inspection rather than adjudicating scholarly intent.