Hybrid Feature Selection for COVID-19 Text Classification using Cuckoo Search Optimization and Mutual Information with DeBERTa

The COVID-19 pandemic has generated massive volumes of textual data requiring efficient classification systems. Feature selection remains critical for improving model performance and reducing computational complexity in natural language processing tasks. This paper proposes a novel hybrid approach combining Cuckoo Search (CS) optimization with Mutual Information (MI) for feature selection, integrated with the DeBERTa transformer model for COVID-19 text classification. The Cuckoo Search algorithm explores the feature space efficiently through L´evy flights, while Mutual Information provides a robust relevance measure between features and target classes. Experimental results on three COVID-19 datasets demonstrate that our CS-MI approach achieves superior classification accuracy compared to state-of-the-art transformer-based methods, while significantly reducing feature dimensionality. The proposed method achieves 94.2% accuracy on Twitter data, 93.5% on news articles, and 95.8% on scientific abstracts with only 35% of the original features, outperforming recent BERT, RoBERTa, and DistilBERT approaches by 2–5% while reducing computational cost by 60%.