<p>The dynamic and multimodal nature of digital media has significantly complicated the measurement of news communication influence. Traditional metrics such as reach and engagement fail to capture the deeper cognitive, emotional, and behavioral responses of diverse audiences. To address this limitation, this study proposes a structured multimodal evaluation framework based on a Scalable Locust Swarm Optimized Multimodal Graph Neural Network (SLS-MGNN). First, multimodal data from the Chinese News Communication Influence Dataset including text, images, audio, and video are preprocessed using lemmatization, resizing, and denoising techniques. Feature extraction is then performed using Bag-of-Words (BoW) for text, Convolutional Neural Networks (CNNs) for visual content, and Short-Time Fourier Transform (STFT) for audio signals. These modality-specific representations are fused at the feature level to construct unified embeddings. A Multimodal Graph Neural Network (MGNN) models relationships among news items, audience clusters, and dissemination channels, where nodes represent entities and edges encode engagement patterns, forwarding behavior, and semantic similarity. The model parameters are optimized using a Scalable Locust Swarm (SLS) strategy to enhance convergence stability and global search capability. Experimental results on a 70/15/15 train–validation–test split demonstrate that the proposed model achieves 99% accuracy and F1-score in multi-class classification of public opinion trends and engagement categories, outperforming baseline models such as SVM, Decision Tree, Random Forest, and TextConvoNet. The framework further provides interpretable insights into cognitive, emotional, and behavioral influence patterns. These findings suggest that SLS-MGNN offers a robust and scalable solution for assessing news communication influence in complex digital ecosystems.</p>