Proposed TopoTemp, a framework combining topological data analysis and sequential modeling, achieving improved community evolution prediction on temporal social networks. Analyzed community structures during politically sensitive periods using Louvain, HDBSCAN, and Ward clustering, with LLM-based labeling showing Ward’s clusters as most interpretable.
Designed a Diffusion-Augmented link prediction framework that combines transformer attention with diffusion processes to better capture relationships and influence in social networks. Outperformed state-of-the-art methods by modeling complex relationships and information flow, with a 20% boost in capturing higher-order dependencies.
Analyzed single-modality (RoBERTa + BEiT) and multimodal (BLIP) transformers for VQA optimizing feature
fusion strategies to enhance accuracy and real-world applicability. BLIP, with an F1 score of 30.49%, demonstrated
better semantic understanding.