Abstract
The rapid spread of misleading information on social media influences public behaviour and complicates crisis communication. Although transformer models such as BERT accurately detect misinformation at the post level, most studies analyse posts in isolation and overlook howmisinformation fluctuates over time or responds to major events. This study addresses that gap by developing an end-to-end analytical workflow that integrates BERT-based classification with temporal aggregation, topic clustering, anomaly detection, and event alignment. The analysis uses 10,700 COVID-19–related tweets (6,420 training, 2,140 validation, and 2,140 testing). Because timestamps were unavailable, synthetic timestamps were assigned using an evenly spaced date range between 1 January 2020 and 1 August 2020, and ISO week numbers were derived to support temporal exploration. A fine-tuned BERT-base model trained with early stopping achieved strong test performance (accuracy = 0.956; F1-scores ≈ 0.95– 0.96). Topic discovery using TF–IDF and K-means (with a five-cluster solution) revealed shifting thematic patterns. For anomaly detection, Isolation Forest and a simple peak-based approach were tested; the latter more reliably identified high-activity weeks. Assigning synthetic event labels to the detected spike weeks enabled structured interpretation of peak periods within a simplified timeline (e.g., early alerts, lockdown communications, and vaccine-trial updates), without implying real-world causality. Rather than treating detection as an end in itself, the workflow reframes classifier outputs as signals for temporal monitoring, narrative interpretation, and contextual prioritisation. Overall, the findings show that combining transformer-based classification with lightweight temporal and event-aware methods provides a clearer picture of how misinformation evolves. The framework is modular, interpretable, and suitable for extension to real temporal data, multimodal contexts, or early-warning applications.
Publication Date
12-2025
Document Type
Thesis
Student Type
Graduate
Degree Name
Professional Studies (MS)
Department, Program, or Center
Graduate Programs & Research
Advisor
Sanjay Modak
Advisor/Committee Member
Ioannis Karamitsos
Recommended Citation
Vijayaraghavan, Vishnu Tejas, "Beyond Detection: A Batch-Based AI Framework for Temporal and Event-Correlated Trend Analysis of Misinformation on Social Media" (2025). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12487
Campus
RIT Dubai
