When: Thursday, 3/6 11 am – 12 pm; Where: Tyler Hall 055
Abstract: This talk explores the advancement of Hypergraph Neural Networks (HyperGNNs) as a powerful extension of Graph Neural Networks (GNNs) to model higher-order relationships in complex systems, particularly in biomedical applications. While traditional GNNs struggle to capture higher-order intricate dependencies, HyperGNNs leverage hypergraphs to encode these complex interactions more effectively. In this talk, Dr. Saifuddin will present his research on developing the SOTA novel HyperGNN architectures, and their application in diverse domains. The first study introduces HyGNN, a hypergraph-based model for drug-drug interaction (DDI) prediction, where drugs are represented as hyperedges and their substructures as nodes. This approach integrates an attention-based hyperedge encoder to improve drug representation learning. The second study extends HyperGNNs to sequence representation learning, developing Seq-HyGAN, which captures higher-order relationships in sequences via common subsequences, significantly enhancing representation learning. Moving beyond applications, the third study presents the Topology-guided Hypergraph Transformer (THTN), which integrates structural and spatial attributes into node representations, overcoming the limitations of existing hypergraph transformers that rely only on semantic features. Moreover, it introduces a structure-aware attention mechanism that identifies both structurally and semantically important hypergraph entities (node and hyperedge) and thus enhances representation learning. Finally, the fourth study introduces HyperGCL, a hypergraph contrastive learning framework that generates multiple hypergraph views to capture both structural and attribute-driven multi-model information, utilizing a learnable augmentation function, view-specific encoders, and network-aware contrastive loss. These contributions collectively demonstrate the robustness and versatility of HyperGNNs, bridging theoretical advancements and real-world applications in biomedicine, drug discovery, and multi-modal learning.
Bio: Dr. Khaled Mohammed Saifuddin is a Postdoctoral Research Associate at BarabasiLab, part of the Center for Complex Network Research at Northeastern University, working under the mentorship of Dr. Albert-László Barabási. Dr. Saifuddin’s research focuses on Network Science and Artificial Intelligence, with particular emphasis on Graph Mining, Geometric Machine Learning (Graph and Hypergraph Neural Networks), and Generative AI for Graphs. He applies these methods to Bioinformatics, Health Data Analytics, and Multi-modal Learning.
Before joining BarabasiLab in August 2024, Dr. Saifuddin earned his Ph.D. in Computer Science from Georgia State University under the supervision of Dr. Esra Akbas, where he was involved in doing research in Graph AIs and their application in diverse domains. His research works have been published in leading conferences such as ICDE, CIKM, ECML PKDD, BigData, DSAA, and others.
Dr. Saifuddin has received multiple research awards, including the Best Poster Presentation at the 2023 Machine Learning for Drug Discovery event organized by the Broad Institute of MIT and Harvard, as well as the 1st Place Dell-Intel Student Award for outstanding use of data science and computing. He has served as an Area Chair for the AI4Science workshop at ICML 2024 and NeurIPS 2023 and as a reviewer for prestigious conferences such as KDD, WWW, IJCNN, etc.