Multisource Information-Based Social Event Detection Using Integrated Deep Neural Architectures and Hybrid Text Representation

Event detection is crucial for timely crisis response, public safety, and combating misinformation by extracting critical signals from real-time social data. The user-generated content on social media platforms has created a critical need for intelligent systems capable of accurately detecting and classifying real-world events across diverse domains. This paper presents a fusion-driven framework for social event detection, leveraging textual data from diverse social media sources and a suite of deep neural architectures to achieve state-of-the-art performance in classifying diverse event categories from noisy, unstructured social media streams. The proposed system collects information from heterogeneous domains, including medical conditions, natural disasters, and sports events, to improve both feature richness and contextual awareness. Our method applies a hybridized deep learning strategy, specifically a customized layered CNN+LSTM architecture. This fusion enables the model to learn spatial and temporal dependencies simultaneously, thus enhancing classification performance. To benchmark effectiveness, we evaluated standalone architectures including CNN, RNN, LSTM, BiLSTM, and BiRNN. Among individual models, LSTM and BiLSTM yielded the top performance, achieving up to 95.3% and 95.8% precision, 95.5% and 96.1% recall, 95.4% and 95.9% accuracy, 95.4% and 96.0% F1-score, 95.6% and 96.2% F1.5, and 95.5% and 96.1% F2-scores, respectively. However, these models still showed variability across event types due to limited contextual generalization from a single-layered perspective. In contrast, the proposed fused CNN and LSTM model recorded promising performance with superior F1-scores on 14 independent classes, achieving 100% on Acne, Earthquake, Landslide, Ringworm, and Vitiligo, 99% on Depression, 98% on Badminton and Tennis, 97% on Cricket, and 96% on FIFA (2022). These results demonstrate the model’s robustness across both high- and low-frequency classes, as well as its ability to generalize across varying linguistic and topical contexts. The collection of data from multiple heterogeneous sources not only mitigates the limitations posed by domain-specific biases and class imbalance but also enriches semantic representation and improves the model’s adaptability to emerging events. This research thus establishes a scalable foundation for intelligent, real-time event classification systems in domains such as public safety, health surveillance, and crisis management.