Benchmarking nonnegative tensor factorization for topic modeling: Coherence, separability, and cost
Iberoamerican Journal of Science Measurement and Communication · 30.03.2026 · 6 · 1-18
DOI: 10.47909/ijsmc.349 ↗Kurz gefasst
Fachartikel in der Zeitschrift Iberoamerican Journal of Science Measurement and Communication, online erschienen am 30.03.2026. Verfasst von Andres Grisales-Aguirre und 2 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0). Gefördert von: Universidad Nacional de Colombia, Sede Manizales,Universidad Nacional de Colombia, Sede Manizales.
Abstract (Originalsprache)
Objective
The objective of this study was to determine whether nonnegative tensor factorization (NTF) could enhance topic modeling (TM) in comparison with conventional matrix-based methodologies (latent semantic indexing/latent semantic analysis, latent Dirichlet allocation [LDA], and nonnegative matrix factorization). Furthermore, it aimed to test the hypothesis that a multiway representation could better preserve higher-order term–document structure than two-dimensional document–term matrices. Design/Methodology/Approach. A third-order document representation was constructed, incorporating term entropy as an informativeness mode in conjunction with local weighting (tf-idf ). The resulting tensor was factorized using an NTF procedure based on alternating least squares with nonnegativity constraints and regularization. The performance of the model was evaluated using several metrics. These included topic coherence Cv, qualitative interpretability diagnostics (such as word clouds, topic-similarity heat maps, and t-SNE), and computational cost (measured in training time). Results/Discussion. In a supervised classification setting, cross-validation yielded strong performance for NTF (F1, precision = 0.98, recall = 0.97, accuracy = 0.93) with high coherence (Cv = 0.87). In an unsupervised benchmark aligned with Singh et al. (2023), a coherence-based topic selection method identified an optimal selection of eight topics. Furthermore, the NTF approach yielded higher average coherence than the reference LDA-based results (0.52 vs. 0.47). These gains, however, were accompanied by an augmented computational cost, as evidenced by the extended training time required (e.g., 39 seconds).
Conclusions
Entropy-augmented tensor representations in combination with NTF yielded more coherent and distinctly separated topics than matrix-based baselines, albeit with elevated computational demands. Moreover, these findings suggested that the proposed NTF approach could be highly valuable for applications such as science mapping studies, academic literature analysis, and systematic reviews. Originality/Value. The study empirically validated a term-entropy–augmented tensor construction for TM and demonstrated that preserving multiway structure could improve topic coherence and separability beyond conventional two-dimensional document–term representations.
Förderung
- Universidad Nacional de Colombia, Sede Manizales,Universidad Nacional de Colombia, Sede Manizales
Zitieren
APA
Grisales-Aguirre, A., Robledo, S., & Eggers, F. (2026). Benchmarking nonnegative tensor factorization for topic modeling: Coherence, separability, and cost. Iberoamerican Journal of Science Measurement and Communication, 6, 1-18. https://doi.org/10.47909/ijsmc.349
Vancouver
Grisales-Aguirre A, Robledo S, Eggers F. Benchmarking nonnegative tensor factorization for topic modeling: Coherence, separability, and cost. Iberoamerican Journal of Science Measurement and Communication. 2026;6:1-18. doi:10.47909/ijsmc.349
BibTeX
@article{GrisalesAguirre2026,
author = {Grisales-Aguirre, Andres and Robledo, Sebastian and Eggers, Fabian},
title = {Benchmarking nonnegative tensor factorization for topic modeling: Coherence, separability, and cost},
journal = {Iberoamerican Journal of Science Measurement and Communication},
year = {2026},
volume = {6},
pages = {1--18},
doi = {10.47909/ijsmc.349}
}
EUVIMED