Mayfly-Optimized Multimodal Dual-Channel Text Classification Model

Abstract

Traditional single-channel convolutional neural networks (CNNs) and recurrent neural networks (RNNs) exhibit limitations in text feature extraction, while deep learning models heavily rely on manual hyperparameter tuning. To address these challenges, we propose a Mayfly Algorithm (MA)-optimized dual-channel neural network for multimodal text classification, which enhances both accuracy and stability. The model employs a hybrid architecture: (1) a sequential input module combining Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTM) networks to capture contextual semantics, with MA-driven hyperparameter optimization; (2) a parallel dual-channel CNN (CNN1 and CNN2) to extract local and global features; and (3) a feature fusion layer integrating multimodal representations. The final classification is performed via a Softmax layer. Experiments on the THUCNews dataset demonstrate a state-of-the-art accuracy of 97.2%, significantly outperforming existing methods (e.g., LSTM: 93.3%, CNN: 93.4%). This validates the model’s superior effectiveness in text classification tasks.