Rumor Detection Model on Social Media Based on Bi-LSTM and Hierarchical Attention Mechanism

The complexity of rumors on social media places higher demands on detection models for multimodal feature fusion and semantic understanding. This study proposes a hierarchical attention multimodal rumor detection model that integrates Bidirectional Long Short-Term Memory network for text feature extraction, convolutional autoencoder for user behavior modeling, and Adaptive Graph Convolutional Network for image analysis. The model uses a hierarchical attention mechanism to achieve cross-modal correlation. In ablation experiments, the model without the Adaptive Graph Convolutional Network module achieves an area under the receiver operating characteristic curve of 0.78, which is 11.4% lower than the complete model. In performance comparison experiments, the proposed model achieves an average detection accuracy of 98%, a response time of 40 ms, and a detection accuracy error reduced to 0.1 after 50 iterations, all significantly better than comparison models. The results show that hierarchical focus and dynamic interaction of multimodal features significantly improve rumor detection accuracy and efficiency. This model is more suitable for lightweight deployment and meets the increasing demand for efficient rumor detection on social media.