The rank of a journal is usually quantified by a single metric like Impact Factor (IF), CiteScore or SCImago Journal Rank (SJR). To avoid bias due to the use of a single metric, a multimetric-based data-centric unsupervised machine learning framework is proposed for journal categorization.
This study integrates six bibliometric indicators as features to map the latent structure of journal impact. The proposed framework incorporates two state-of-the-art dimensionality reduction algorithms (PCA and t-SNE), which also contribute to feature engineering. Using PCA, the correlated metrics are transformed into reduced dimensions (2D/3D) that serve as input for k-means clustering. The optimal clusters (k = 4), aligned with IF, Scopus and SJR quartiles, are validated through internal and external evaluation measures.
Multimetric-based clustering of 4,146 computer science (CS) journals is compared with established single-metric approaches (SJR, IF and Scopus quartiles) and with HJRS's proprietary multimetric approach, adopted in Pakistan. The average metric-based results provide interpretable outcomes against SJR, IF and Scopus quartiles (single-metric) as well as HJRS (multi-metric). The proposed journal categorization method is fully automated and offers improved classification of journals compared to well-known traditional metrics.
The contribution of this study is evident in the proposed framework for categorizing journals using multiple metrics. The novelty of our work lies in the integration and domain-specific adaptation of unsupervised machine learning to categorize CS journals.
The peer review history for this article is available at: Link to the website
