<![CDATA[Deep neural architectures have demonstrated substantial capability for handling temporal and sequential data; however, most recurrent-based models, such as LSTM, BiLSTM, GRU, and BiGRU, remain computationally expensive and prone to overfitting. This study proposes and evaluates the EMOGRAM-CNN model, a convolutional neural architecture enhanced with Gram-matrix feature correlation, to improve feature representation in temporal classification tasks. Model performance was compared with conventional CNNs and recurrent architectures on a balanced six-class dataset comprising 17,967 samples. Experimental results show that EMOGRAM-CNN achieved the highest classification accuracy of 94.48%, outperforming CNN (94.00%), GRU (92.00%), BiGRU (91.00%), BiLSTM (91.00%), and LSTM (90.00%). The model converged faster, with smoother loss behavior and lower validation error, indicating superior stability and generalization. The Gram-based correlation layer effectively preserved second-order dependencies across feature maps, enabling the network to capture both local and global temporal relationships without recurrent connections. These findings confirm that EMOGRAM-CNN offers a robust, computationally efficient alternative to recurrent deep networks for sequence classification.]]>
Copyrights © 2026
