Experimental Evaluation of Deep Learning Architectures for Large-Scale Data Processing and Analytics

Abstract

Deep learning has emerged as a powerful paradigm for large-scale data processing and analytics, enabling automated feature extraction and high-performance modeling across diverse domains such as healthcare, finance, and smart infrastructure. This research presents an experimental evaluation of major deep learning architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Transformer models, and Graph Neural Networks (GNNs), in large-scale data environments. The study investigates their performance, scalability, computational efficiency, and adaptability across heterogeneous datasets. A systematic experimental framework is designed by integrating distributed computing platforms and parallel processing techniques to simulate real-world big data conditions. The evaluation highlights that Transformer-based architectures outperform traditional sequential models in large-scale text analytics due to their parallel processing capability, while CNNs remain highly efficient for spatial data processing. Furthermore, the study identifies trade-offs between model complexity and computational cost, emphasizing the importance of hybrid and optimized architectures for scalable analytics. The findings contribute to a deeper understanding of architecture selection and optimization strategies in big data ecosystems, providing insights for future research in scalable and efficient deep learning systems.