SciPred — Predatory Journal Detection
NLP pipeline to detect predatory journals using SciBERT embeddings
Overview
SciPred is an end-to-end text classification pipeline I developed to detect predatory journals by analyzing article and journal text features using SciBERT embeddings and a lightweight MLP classifier.
Approach
- Collected and cleaned a dataset of journal and article metadata and descriptions; used web-scraping with BeautifulSoup to assemble training data.
- Embedded textual content with SciBERT and trained an MLP classifier to distinguish predatory outlets from legitimate journals.
- Applied careful preprocessing and class-balancing to improve model stability.
Results
- Achieved an F1 score of 0.92 on held-out data, outperforming baseline text-only models by ~8%.
Stack
- Python, HuggingFace Transformers (SciBERT), PyTorch, scikit-learn, BeautifulSoup.