The rapid expansion of large-scale unstructured data has introduced substantial challenges for effective pattern discovery, particularly in achieving a balance between high-capacity representation learning and interpretability. Although deep learning techniques have demonstrated exceptional capability in extracting rich latent representations from unstructured data, they often lack explicit mechanisms for generating interpretable knowledge patterns required for transparent and explainable analytical systems. To address this limitation, the present study proposes a Hybrid Deep Learning–Data Mining (H-DLDM) framework that systematically integrates self-supervised representation learning with explicit data mining methodologies. The proposed framework was evaluated using three publicly available datasets representing textual, visual, and multimodal data domains. Experimental findings demonstrate that the H-DLDM framework generates semantically cohesive latent representations, supports hierarchical and relational pattern discovery, and achieves a balanced trade-off among pattern quality, interpretability, computational efficiency, and scalability. Comparative analysis further reveals that the proposed framework consistently outperforms both deep learning–only and conventional data mining baseline approaches by producing more stable, interpretable, and semantically meaningful patterns without compromising large-scale analytical performance. Overall, the findings highlight the potential of hybrid analytical frameworks to transform deep latent representations into explicit and actionable knowledge structures. The proposed approach therefore contributes to the advancement of explainable, scalable, and knowledge-centric data mining systems for complex unstructured data environments.
Shahid et al. (Sat,) studied this question.