EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A multimodal visual framework for predatory journal detection: dataset construction, hand-engineered visual features, and cross-modality fusion

Abubaker Mohamed, Kevin A. McCormick, Moshiur Farazi

Open Computer Science · 23.01.2026 · 16 (1)

DOI: 10.1515/comp-2025-0094 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Open Computer Science, online erschienen am 17.11.2026. Verfasst von Abubaker Mohamed und 2 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Predatory academic journals mimic the visual and structural conventions of legitimate peer-reviewed venues while operating without meaningful editorial oversight, undermining scholarly integrity at a global scale. Existing defences depend on human-curated blacklists that are inherently reactive and do not scale. We integrate three independently trained and individually well-established classifiers into a late-fusion framework for this domain: a Random Forest on structured website features, a calibrated Gradient Boosting classifier on term-frequency–inverse-document-frequency (TF-IDF) representations of optical-character-recognition (OCR) text, and a visual branch in which a fine-tuned EfficientNet-B0 supplies the fused probability, accompanied by an interpretable Gradient Boosting classifier on 79 hand-engineered visual features extracted from homepage screenshots. Our contribution is not a new model but the construction, integration, and leakage-controlled evaluation of this pipeline in a new application domain. Three elements are emphasised. First, we build and openly release a multimodal dataset of journal websites with paired structured metadata, OCR-extracted text, and screenshots-derived visual features, assembled through a documented scraping and OCR pipeline. Second, we organise the visual features into eight perceptual modules (Layout, Color, Typography, Whitespace, Complexity, Trust, Ads, Professional) that provide per-feature interpretability absent from learned image encoders. Third, because legitimate labels derive from the Directory of Open Access Journals (DOAJ), we explicitly quantify and control the resulting label leakage by reporting every result in both a DOAJ-aware and a DOAJ-blind regime. Under five-fold cross-validation on 1,101 journals, the structured branch alone reaches AUROC 0.997 in the DOAJ-aware regime but 0.908 once the indexing feature is removed. A learned stacking ensemble attains an F1 of 99.1 % in the DOAJ-aware regime and, in the deployment-oriented DOAJ-blind regime, an F1 of 93.5 % (93.3 % under a strictly nested stacking protocol), which significantly exceeds every individual modality.

Zitieren

APA

Mohamed, A., McCormick, K. A., & Farazi, M. (2026). A multimodal visual framework for predatory journal detection: dataset construction, hand-engineered visual features, and cross-modality fusion. Open Computer Science, 16(1). https://doi.org/10.1515/comp-2025-0094

Vancouver

Mohamed A, McCormick KA, Farazi M. A multimodal visual framework for predatory journal detection: dataset construction, hand-engineered visual features, and cross-modality fusion. Open Computer Science. 2026;16(1). doi:10.1515/comp-2025-0094

BibTeX
@article{Mohamed2026,
  author = {Mohamed, Abubaker and McCormick, Kevin A. and Farazi, Moshiur},
  title = {A multimodal visual framework for predatory journal detection: dataset construction, hand-engineered visual features, and cross-modality fusion},
  journal = {Open Computer Science},
  year = {2026},
  volume = {16},
  number = {1},
  doi = {10.1515/comp-2025-0094}
}
Kostenlos starten →