Preprint
Concept Paper

This version is not peer-reviewed.

Composite Edge Weighting for Leiden Community Detection in a GraphRAG-Style Knowledge Graph: Frequency, Relationship Strength and Semantic Similarity

Submitted:

12 November 2026

Posted:

13 November 2026

You are already at the latest version

Abstract
Graph-based Retrieval-Augmented Generation (GraphRAG) organizes a corpus as a knowledge graph of entities and relationships, partitions it into communities with the Leiden algorithm, and summarizes each community to answer corpus-level questions. The partition depends on the edge weights supplied to the clustering step, yet the weighting scheme is usually a single heuristic, namely the sum of LLM-assigned relationship strengths. This paper presents a reproducible experimental framework that compares edge-weighting signals for Leiden community detection. Three normalized features are derived for each edge: a log-normalized distinct-evidence frequency (F), a normalized mean LLM relationship strength (S), and an embedding-based similarity of the endpoint entity descriptions (C). Nine weighting variants are evaluated: a sum-of-strengths baseline, three single-feature weights, a frequency-plus-strength weight, the proposed composite W = 0.4F + 0.4S + 0.2C, and three leave-one-component-out ablations. The pipeline uses a locally hosted LLM (Qwen2.5-3B via Ollama) for evidence-grounded extraction, sentence-transformer embeddings (all- MiniLM-L6-v2), and weighted Leiden clustering with controlled seeds, and it reports NMI, ARI, seed stability, singleton share, connectivity and weighted modularity. A Streamlit dashboard exposes the same weighting logic for interactive, query-time exploration. On a small synthetic institutional-policy corpus (16 graph nodes, 14 undirected edges), all nine variants yielded the identical partition (3 communities, NMI = 0.901, ARI = 0.789, seed-stability ARI = 1.0); only the weighted modularity diagnostic differed, within a range of 0.0058. We trace this result to a lack of signal diversity in the graph: every edge had exactly one evidence instance, so F was constant, and S and C varied only narrowly. We therefore report no evidence that the composite weighting improves community quality at this scale, and we discuss the conditions under which the comparison becomes informative.
Keywords: 
;  ;  ;  ;  ;  ;  ;  

1. Introduction

Large language models (LLMs) answer questions well when the needed facts are in their context, but they struggle with global, corpus-level questions such as “what are the main themes in this collection?”. Retrieval-augmented generation (RAG) [9,10] grounds generation in retrieved passages, but vector retrieval of a few chunks cannot summarize an entire corpus. GraphRAG addresses this by using an LLM to extract a knowledge graph, partitioning the graph into hierarchical communities with the Leiden algorithm, and generating summaries per community that are later combined into a global answer [1]. Subsequent systems and surveys have extended the idea of graph-structured retrieval [11,12,13,14].
The quality of the community summaries depends on the quality of the partition, and the partition depends on the graph’s edge weights. In the reference formulation, an edge weight is aggregated from the strength ratings that an LLM assigns to extracted relationships [1]. This is a single signal. It conflates how often a relationship is stated, how directly a sentence expresses it, and how semantically close its endpoints are. Weighted-network analysis has long shown that how weights are defined changes the structure that is detected [5,19,20], and the resolution and quality function of modularity-based methods strongly shape the communities they return [7,8].
This work asks a focused question: does replacing the single strength-sum weight with a composite of frequency, strength and semantic similarity change the communities found by Leiden, and does it do so in a way that agrees better with known topics? We make four contributions.
  • A modular, cached and seed-controlled pipeline that extracts an evidence-validated knowledge graph with a local LLM and evaluates nine edge-weighting variants on identical graphs.
  • A transparent feature design (F, S, C) in which every feature is bounded in [0, 1], and a set of ablations that isolate each component.
  • An evaluation protocol combining external agreement (NMI, ARI), internal diagnostics (modularity, singleton share, components), and partition stability across seeds.
  • An honest negative result on a small corpus, with a diagnosis of why the variants are indistinguishable and what the data must contain for the comparison to be meaningful.

3. Methodology

3.1. Overview

Figure 1 summarizes the pipeline: corpus cleaning and section-level chunking, LLM extraction with evidence validation, entity resolution, evidence aggregation into edges, feature computation, weighted Leiden clustering, and metric computation. Extraction outputs are cached per chunk (keyed by content hash, model and prompt version) so that every weighting variant is evaluated on the same frozen graph.

3.2. Corpus

The corpus consists of four synthetic institutional-policy files covering examination regulations, faculty and administration, hostel rules, and medical leave. The files contain deliberate synthetic expansions and limitation notes. A cleaning step removes document-number lines, synthetic-data notes and limitation statements so that the extractor sees only policy sentences. Each file is split at Document: boundaries into section-level documents, and exact duplicate sections are discarded by SHA-256 hash. Because the text is synthetic, it does not establish real institutional policy.

3.3. Evidence-Grounded Extraction

Each section is sent to a locally hosted Qwen2.5-3B model [28] through Ollama, with temperature 0, a context window of 8192 tokens and a JSON-only output contract. The prompt requests entities (short canonical name and description) and relationships comprising source, target, a normalized relation phrase, a description, one evidence sentence, and an integer strength from 1 to 10. The strength rubric anchors 1-2 as incidental, 5-6 as moderate and explicit, and 9-10 as explicitly defining, and it states that strength must reflect how directly a sentence expresses the relation, not topic popularity.
To limit hallucinated relations, each evidence string is matched against the sentences of its source section. A sentence is accepted if the normalized evidence is contained in a source sentence (or vice versa), or if the best token-level Jaccard overlap is at least 0.5. The verbatim source sentence replaces the model’s string. Relations without a matching sentence, with self-loops after normalization, or with over-long names are discarded. Failed or malformed JSON responses are retried up to three times, and empty extractions are never cached.

3.4. Entity Resolution

Entity names are normalized by lowercasing, stripping punctuation, collapsing whitespace and singularizing the final word (for words longer than three characters that do not end in “ss”, “us” or “is”). Entities sharing a normalized key are merged and the longest description is retained. This is deliberately conservative and does not resolve aliases such as “student” and “students” when they are handled as distinct keys.

3.5. Edge Aggregation and Features

Relationship instances are grouped by (source key, target key, relation type). Distinct evidence is counted by evidence identifier, so duplicate sentences do not inflate counts. For an edge e with n_e distinct evidence items and mean strength s̄_e:
F_e = log(1 + n_e) / log(1 + n_max) (F_e = 1 for all edges if n_max <= 1)S_e = (s̄_e - 1) / 9C_e = (cos(v_src, v_tgt) + 1) / 2
where v is the L2-normalized all-MiniLM-L6-v2 embedding [22] of the string “name: description” for each endpoint. F is a logarithmically damped, corpus-normalized count of evidence, S rescales the LLM rating to [0, 1], and C is an affine map of cosine similarity. C measures the similarity of entity descriptions, not the correctness of the relation.

3.6. Weighting Variants

ID Variant Weight
A Sum of strengths (baseline) Σ strength over distinct evidence
B Frequency only F
C Strength only S
D Similarity only C
E Frequency + strength 0.5F + 0.5S
F Composite (proposed) 0.4F + 0.4S + 0.2C
G Ablation: no frequency (0.4S + 0.2C) / 0.6
H Ablation: no strength (0.4F + 0.2C) / 0.6
I Ablation: no similarity 0.5F + 0.5S
Variant A follows the sum-of-strengths aggregation of the project methodology and has not been verified against a pinned release of the official GraphRAG code; we call it a baseline in the style of [1], not a reproduction. Variants E and I are mathematically identical by construction. All weights are clipped below at 10−9.

3.7. Graph construction and Clustering

The weighted graph is undirected. Directed and opposite-direction edges between the same pair of nodes are collapsed and their weights summed, and self-loops are dropped. This choice is fixed across variants. Communities are found with the Leiden algorithm [2] using the RB-configuration vertex partition (modularity with a resolution parameter) at resolution 1.0 and a controlled seed, implemented with python-igraph and leidenalg [32]. Leiden cluster labels are renumbered for reproducibility.

3.8. Evaluation Metrics

  • External agreement: NMI [25,36] and ARI [23,24] between Leiden communities and topic labels, computed on the labelled nodes only (via scikit-learn [30]). Labels are resolved by exact normalized match and then by fuzzy match with difflib cutoff 0.88. Entities with conflicting labels are ignored.
  • Stability: mean ARI between the base partition and partitions from further seeds, and an adjacent-seed ARI across repeats.
  • Structure: number of communities, singleton share, weakly connected components, isolated nodes, and minimum, maximum and mean community size.
  • Internal diagnostic: weighted modularity of the returned partition. Because each variant changes the weights, modularity values are not strictly comparable across variants and are used only as a diagnostic.

3.9. Experimental Protocol

The runner supports corpus fractions and repeated document-level sampling with seed = base seed + 10000 × fraction index + repeat. The reported run used the full corpus (fraction 1.0), five repeats, base seed 42 (seeds 42-46), and resolution 1.0, for 9 variants × 5 repeats = 45 clustering runs. Because the full corpus was used, the extracted graph was identical across repeats and only the Leiden seed varied.

3.10. Interactive Dashboard

A Streamlit application implements a query-time version of the pipeline. For a user query, it retrieves the top-k (default 4) section-level documents with TF-IDF vectors (unigrams and bigrams, English stop words) and cosine similarity [33,34]. It then extracts a relationship graph from the retrieved text with a local Ollama model, computes baseline and composite weights, renders the graph with Plotly, and generates an answer grounded in the retrieved passages and graph triples. The user can toggle between baseline and composite weighting. In the dashboard, the semantic term C is a TF-IDF cosine between endpoint names, not an embedding similarity, and the composite uses raw mean strength rather than the normalized S (Section 7). The dashboard is therefore an exploration tool and is not the source of the quantitative results.

3.11. Implementation and Reproducibility

The system is implemented in Python with pandas, scikit-learn [30], NetworkX [31], igraph/leidenalg, sentence-transformers [21], Plotly and Streamlit. The configuration (model names, URL, seed, repeats, resolution, formulas, prompt version) is written to a JSON file with each run, and edge-level provenance (evidence identifiers, texts and source paths) is saved alongside the weights.

4. Experimental Setup

Item Setting
Extractor Qwen2.5-3B (qwen2.5:3b) via Ollama, temperature 0, 8192-token context
Embedding model sentence-transformers/all-MiniLM-L6-v2 (enabled)
Clustering Leiden, RB-configuration partition, resolution 1.0
Seeds / repeats Base seed 42, 5 repeats (seeds 42-46)
Source files 4 synthetic policy files, section-level chunks
Extracted entities / relation instances 22 / 19
Aggregated directed edges 19
Graph used for clustering 16 nodes, 14 undirected edges
Labelled nodes per run 8 of 16 (topic labels from a hand-made file)
Of the 22 extracted entity records, 6 did not participate in any retained relationship and are therefore not graph nodes. The 19 directed edges collapsed into 14 undirected node pairs.

5. Results

5.1. Edge Feature Distributions

All 19 aggregated edges were supported by exactly one distinct evidence sentence. Consequently n_max = 1 and F = 1 for every edge, so F carries no discriminative information in this run. LLM strength ratings ranged from 7 to 9 (mean 8.05, standard deviation 0.62), giving S between 0.667 and 0.889. The unit-scaled similarity C ranged from 0.536 to 0.831 (mean 0.714, standard deviation 0.102). The composite weight therefore varied only through S and C, and by construction C contributes at most 20% of the weight.

5.2. Partition Quality Across Weighting Variants

All nine variants returned the same partition in every run: three communities of sizes between 3 and 8 (mean 5.33), no singletons and no isolated nodes. The adjacent-seed ARI between repeats was 1.0 in the per-run records inspected, indicating that the partition does not depend on the Leiden seed at this scale.
Table 1. Full-corpus results, mean over 5 seeds (standard deviation of NMI, ARI and all structure metrics is 0, apart from NMI at ~1.2 × 10−16 numerical noise).
Table 1. Full-corpus results, mean over 5 seeds (standard deviation of NMI, ARI and all structure metrics is 0, apart from NMI at ~1.2 × 10−16 numerical noise).
Variant NMI ARI Communities Singleton share Components Weighted modularity
A Sum of strengths 0.9007 0.7895 3 0 3 0.6327
B Frequency only 0.9007 0.7895 3 0 3 0.6371
C Strength only 0.9007 0.7895 3 0 3 0.6321
D Similarity only 0.9007 0.7895 3 0 3 0.6379
E Frequency + strength 0.9007 0.7895 3 0 3 0.6350
F Composite (proposed) 0.9007 0.7895 3 0 3 0.6357
G No frequency 0.9007 0.7895 3 0 3 0.6344
H No strength 0.9007 0.7895 3 0 3 0.6376
I No similarity 0.9007 0.7895 3 0 3 0.6350

5.3. Recovered Communities

The three communities correspond to semantically coherent groups:
  • Examinations: controller of examinations, examination office, students.
  • Academic administration and medical leave: student, Dean of Academic Affairs, Dean, department head, department office, registrar, course instructor, medical practitioner.
  • Hostel: hostel residents, hostel office, warden, mess, student leave process.
The reference labels distinguish four topics (examinations, faculty administration, hostel, medical leave), but the graph merges the faculty-administration and medical-leave entities into one community, because the medical-leave policy is expressed through the same actors (Dean, department office, course instructor, student). This merge limits NMI to 0.90 and ARI to 0.79 on the eight labelled nodes irrespective of the weights.

5.4. Modularity Diagnostic

Weighted modularity ranged from 0.6321 (strength only) to 0.6379 (similarity only), a spread of 0.0058. Since the weights differ across variants, these numbers describe each variant’s own weighted graph and should not be read as a ranking of partition quality; the partitions themselves are identical.

6. Discussion

Why the variants coincide. The identical outcome follows directly from the data. With F constant, variants B, E, G, H and I reduce to monotone functions of S, C or a constant, and with S confined to 7-9 and C to a moderate band, the relative ordering of edge weights barely changes the modularity-optimal partition on a graph whose three groups are already separated by sparse cross-group links. In such a graph, the cluster structure is dominated by topology, and any positive weighting recovers it. The Leiden algorithm is also deterministic enough on 16 nodes that seeds do not matter.
What the result does and does not show. The experiment shows that the pipeline runs end to end, is reproducible, and produces topic-coherent communities with all nine weightings. It does not show that the composite weight is better or worse than the baseline. A tie on a tiny graph is uninformative about larger corpora where frequency varies widely, where edges repeat across many documents, and where cross-topic bridges compete with intra-topic links.
Role of each feature. F can only help when the same relation is stated in several distinct places, which requires a larger and less repetitive corpus than the one used. S depends on the calibration of a 3B-parameter extractor, whose ratings concentrated in a narrow range, and the rubric’s request for variation was largely not reflected. C may down-weight legitimate cross-topic relations, such as a student relating to the Dean, because endpoints from different topics are semantically distant, which is the opposite of the bridging role that weak ties play [18].
Extraction quality. Inspection of the relationship table shows that the small extractor sometimes assigns an incorrect endpoint, for example attributing an examination-office statement to “students” as the source, or linking “department head” to “department office” for a statement about records. Such errors alter the graph structure itself and likely matter more than the weighting. The evidence check confirms that the evidence sentence exists in the source, but not that the extracted source and target are correct.

7. Limitations and Threats to Validity

  • Scale and synthetic data. Four synthetic files and 19 relation instances are far too small for statistical conclusions. No significance tests are meaningful here, and the repeats vary only the Leiden seed, not the graph.
  • Labels. The topic labels are hand-made, only 8 of 16 nodes were labelled per run, and agreement with such labels is not ground truth [26].
  • Extractor. One small model with a single prompt was used. Strength ratings were not audited against human judgment, and endpoint errors occur.
  • Entity resolution. Name-level normalization does not merge aliases (for example “Dean” and “Dean of Academic Affairs”, or “student” and “students”), which fragments hubs.
  • Fixed design choices. The coefficients 0.4/0.4/0.2, undirected collapsing, summed parallel weights, and resolution 1.0 were not tuned or tested for sensitivity.
  • Baseline fidelity. Variant A is not verified against a pinned official GraphRAG release.
  • Dashboard consistency. In the dashboard, C is a TF-IDF similarity of endpoint names and S is not rescaled to [0, 1], so the composite there mixes a 1-10 strength with [0, 1] terms and is dominated by strength. Its behavior is not equivalent to the experiment pipeline and should be aligned before it is used for comparisons. The dashboard also defaults to a different LLM (Llama 3.1 8B [29]) than the experiments.
  • Downstream quality. Community summaries and end-to-end answer quality were not evaluated.

8. Conclusion and Future Work

We presented a reproducible framework for studying edge weights in GraphRAG community detection, with three bounded features, nine weighting variants, and a metric suite combining external agreement, structure and stability. On a small synthetic policy corpus, every variant produced the same three topic-coherent communities (NMI 0.901, ARI 0.789), and the explanation lies in the data: frequency was constant and the other signals varied narrowly. The study therefore establishes a working baseline pipeline, not evidence of an advantage for the composite weight.
Future work should (i) scale to a larger deduplicated corpus where evidence frequency varies, (ii) use a stronger extractor with self-consistency checks and audit samples of entities, relations and strengths, (iii) add alias-aware entity resolution, (iv) build a controlled synthetic benchmark with known topics and weak and strong cross-topic bridges, (v) run coefficient and resolution sensitivity analyses with proper statistics over varied graphs, (vi) verify the baseline against a pinned official GraphRAG release, (vii) align the dashboard features with the experimental ones, and (viii) evaluate community-summary quality and downstream answer accuracy.

Reproducibility Statement

The experiment is run by run_experiment.py with the configuration saved to run_config.json (extractor qwen2.5:3b, embedding model all-MiniLM-L6-v2, 5 repeats, resolution 1.0, seed 42). Outputs comprise edge_weights.csv, experiment_results.csv, summary_statistics.csv and community_memberships.csv. The extraction cache is keyed by section hash, model and prompt version (v2).

References

  1. Edge, D.; Trinh, H.; Cheng, N.; Bradley, J.; Chao, A.; Mody, A.; Truitt, S.; Larson, J. From Local to Global: A Graph RAG Approach to Query-Focused Summarization. arXiv 2024, arXiv:2404.16130. [Google Scholar]
  2. Traag, V. A.; Waltman, L.; van Eck, N. J. From Louvain to Leiden: guaranteeing well-connected communities. Sci. Rep. 2019, 9, 5233. [Google Scholar] [CrossRef] [PubMed]
  3. Blondel, V. D.; Guillaume, J.-L.; Lambiotte, R.; Lefebvre, E. Fast unfolding of communities in large networks. J. Stat. Mech. Theory Exp. 2008, 2008, P10008. [Google Scholar] [CrossRef]
  4. Newman, M. E. J.; Girvan, M. Finding and evaluating community structure in networks. Phys. Rev. E 2004, 69, 026113. [Google Scholar] [CrossRef] [PubMed]
  5. Newman, M. E. J. Analysis of weighted networks. Phys. Rev. E 2004, 70, 056131. [Google Scholar] [CrossRef] [PubMed]
  6. Fortunato, S. Community detection in graphs. Phys. Rep. 2010, 486, 75–174. [Google Scholar] [CrossRef]
  7. Fortunato, S.; Barthélemy, M. Resolution limit in community detection. Proc. Natl. Acad. Sci. 2007, 104, 36–41. [Google Scholar] [CrossRef] [PubMed]
  8. Reichardt, J.; Bornholdt, S. Statistical mechanics of community detection. Phys. Rev. E 2006, 74, 016110. [Google Scholar] [CrossRef] [PubMed]
  9. Lewis, P.; et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems (NeurIPS) 2020, 33, 9459–9474. [Google Scholar]
  10. Gao, Y.; et al. Retrieval-augmented generation for large language models: A survey. arXiv 2023, arXiv:2312.10997. [Google Scholar]
  11. Peng, B.; et al. Graph retrieval-augmented generation: A survey. arXiv 2024, arXiv:2408.08921. [Google Scholar]
  12. Han, H.; et al. Retrieval-augmented generation with graphs (GraphRAG). arXiv 2025, arXiv:2501.00309. [Google Scholar]
  13. Guo, Z.; Xia, L.; Yu, Y.; Ao, T.; Huang, C. LightRAG: Simple and fast retrieval-augmented generation. arXiv 2024, arXiv:2410.05779. [Google Scholar]
  14. Sarthi, P.; Abdullah, S.; Tuli, A.; Khanna, S.; Goldie, A.; Manning, C. D. RAPTOR: Recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations (ICLR); 2024. [Google Scholar]
  15. Pan, S.; Luo, L.; Wang, Y.; Chen, C.; Wang, J.; Wu, X. Unifying large language models and knowledge graphs: A roadmap. IEEE Trans. Knowl. Data Eng. 2024, 36, 3580–3599. [Google Scholar] [CrossRef]
  16. Hogan, et al. Knowledge graphs. ACM Comput. Surv. 2021, 54, 71. [Google Scholar]
  17. Ji, S.; Pan, S.; Cambria, E.; Marttinen, P.; Yu, P. S. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Trans. Neural Netw. Learn. Syst. 2022, 33, 494–514. [Google Scholar] [CrossRef] [PubMed]
  18. Granovetter, M. S. The strength of weak ties. Am. J. Sociol. 1973, 78, 1360–1380. [Google Scholar] [CrossRef] [PubMed]
  19. Barrat, A.; Barthélemy, M.; Pastor-Satorras, R.; Vespignani, A. The architecture of complex weighted networks. Proc. Natl. Acad. Sci. 2004, 101, 3747–3752. [Google Scholar] [CrossRef] [PubMed]
  20. Opsahl, T.; Agneessens, F.; Skvoretz, J. Node centrality in weighted networks: Generalizing degree and shortest paths. Soc. Netw. 2010, 32, 245–251. [Google Scholar] [CrossRef]
  21. Reimers, N.; Gurevych, I. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proc. EMNLP-IJCNLP; 2019; pp. 3982–3992. [Google Scholar]
  22. Wang, W.; Wei, F.; Dong, L.; Bao, H.; Yang, N.; Zhou, M. MiniLM: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems (NeurIPS) 2020, 33, 5776–5788. [Google Scholar]
  23. Hubert, L.; Arabie, P. Comparing partitions. J. Classif. 1985, 2, 193–218. [Google Scholar] [CrossRef]
  24. Vinh, N. X.; Epps, J.; Bailey, J. Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance. J. Mach. Learn. Res. 2010, 11, 2837–2854. [Google Scholar]
  25. Danon, L.; Díaz-Guilera, A.; Duch, J.; Arenas, A. Comparing community structure identification. J. Stat. Mech. Theory Exp. 2005, 2005, P09008. [Google Scholar] [CrossRef]
  26. Peel, L.; Larremore, D. B.; Clauset, A. The ground truth about metadata and community detection in networks. Sci. Adv. 2017, 3, e1602548. [Google Scholar] [CrossRef] [PubMed]
  27. Fortunato, S.; Hric, D. Community detection in networks: A user guide. Phys. Rep. 2016, 659, 1–44. [Google Scholar] [CrossRef]
  28. Yang, et al. Qwen2.5 technical report. arXiv 2024, arXiv:2412.15115. [Google Scholar]
  29. Dubey, et al. The Llama 3 herd of models. arXiv 2024, arXiv:2407.21783. [Google Scholar]
  30. Pedregosa, F.; et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 2011, 12, 2825–2830. [Google Scholar]
  31. Hagberg, A.; Schult, D. A.; Swart, P. J. Exploring network structure, dynamics, and function using NetworkX. In Proc. 7th Python in Science Conference (SciPy); 2008; pp. 11–15. [Google Scholar]
  32. Csárdi, G.; Nepusz, T. The igraph software package for complex network research. InterJournal Complex Syst. 2006, 1695. [Google Scholar]
  33. Spärck Jones, K. A statistical interpretation of term specificity and its application in retrieval. J. Doc. 1972, 28, 11–21. [Google Scholar] [CrossRef]
  34. Salton, G.; Buckley, C. Term-weighting approaches in automatic text retrieval. Inf. Process. Manag. 1988, 24, 513–523. [Google Scholar] [CrossRef]
  35. Newman, M. E. J. Modularity and community structure in networks. Proc. Natl. Acad. Sci. 2006, 103, 8577–8582. [Google Scholar] [CrossRef] [PubMed]
  36. Strehl, A.; Ghosh, J. Cluster ensembles: A knowledge reuse framework for combining multiple partitions. J. Mach. Learn. Res. 2002, 3, 583–617. [Google Scholar]
Figure 1. Pipeline: Corpus → clean + section split → LLM extraction (entities, relations, evidence, strength) → evidence validation → entity resolution → edge aggregation → F, S, C features → nine weight variants → Leiden → NMI / ARI / stability / structure metrics.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.