Meta-learning contrastive fusion intelligence for domain-adaptive content categorization using adaptive DNN

Abstract

The rapid growth of heterogeneous digital text sources, including social media, news streams, and scientific literature, poses significant challenges to traditional content categorization models due to domain shift, vocabulary drift, and semantic inconsistencies. Existing deep learning approaches often rely on domain-specific patterns and static representations, resulting in degraded performance when applied to unseen or cross-domain data. Moreover, these methods lack scalability in real-world scenarios characterized by domain drift and limited labeled data. To address these challenges, this study proposes a meta-learning contrastive fusion intelligence (MCFI) framework for domain-adaptive content categorization. The framework integrates domain context-aware normalization (DCAN) for robust preprocessing, binary particle swarm optimization (BPSO) for selecting domain-invariant features, and a hybrid architecture combining contrastive learning-enhanced bidirectional encoder representations from transformers (CL-BERT) with capsule network-enhanced transformer (CapsTrans). A meta-learning strategy is employed to learn domain-invariant representations and enable rapid adaptation to new domains, while contrastive learning enhances inter-domain separability. A domain-adaptive decision layer further refines feature contributions dynamically. Experimental results on multiple benchmark datasets demonstrate that the proposed MCFI framework consistently outperforms state-of-the-art methods in terms of accuracy, F1-score, and generalization, providing a scalable and effective solution for cross-domain text classification.