Research

Topic: Expanding the Horizons of Hypergraph Neural Networks — Advanced Methodologies

System architecture of HyperGCL. After constructing three different hypergraph views from the input graph and node attributes, we exploit a learnable view augmentation technique to generate adaptive views. View-specific encoders are used to learn each view and finally, a network-aware contrastive loss is used with a supervised loss to train the model.

HyperGCL: Multi-Modal Graph Contrastive Learning via Learnable Hypergraph Views

📘 IJCNN'25  ·  Hypergraph GCLMultimodal Learning Adaptive AugmentationNetwork-aware CL

Overview. HyperGCL constructs three complementary hypergraph views from structure and attributes, then applies a learnable topology-aware augmentation. A network-aware contrastive loss aligns modalities while preserving important relations and filtering noise.

Topology-guided Hypergraph Transformer Network, THTN, consists of a) Structural and Spatial Encoding that enriches initial node representation via learnable structural and spatial node features and b) Structure-Aware Attention that enables the integration of the structural importance of nodes and hyperedges into regular attribute-based semantic attention.

THTN: Topology-Guided Hypergraph Transformer Network

📘 Stanford Graph Learning Workshop'24 · PRICAI'25  ·  Graph Transformer Structure-Aware Attention

Overview. THTN, a topology-guided hypergraph transformer that captures higher-order relations between nodes in a network by representing subgraphs as hyperedges. Then, it introduces a structure-aware attention mechanism that jointly considers semantic and topological cues to identify critical nodes and hyperedges. In addition, a structural–spatial encoding module integrates graph topology and spatial context into node embeddings, enhancing the model’s ability to capture both local and global dependencies.

Topic: Tackling Oversmoothing in GNN

Truss decomposition-based edge pruning framework for graph sparsification to tackle oversmoothing problem

TGS: Tackling Oversmoothing in GNN via Graph Sparsification

📘 ECML PKDD'24  ·  Truss-based Pruning Expressivity

Overview. Truss-based graph sparsification (TGS) model to mitigate the oversmoothing problem in deep GNNs. By pruning redundant edges from dense graph regions, TGS preserves essential topological structures while preventing excessive feature mixing.

Topic: Revolutionizing Drug Discovery with Novel Graph/Hypergraph Learning Approaches

System architecture of the proposed method. The First step is to construct a hypergraph network of drugs where each drug is a hyperedge, and frequent chemical substructures of drugs are the nodes. The second step is to design a hypergraph neural network (HyGNN) model with an attention-based encoder for hyperedge (drug) representation learning and decoder for DDI learning

HyGNN: Hypergraph Neural Network for Drug Representation for DDI Prediction

📘 ICDE'23  ·  Drug Discovery HyperGNN

Overview. HyGNN- a novel hypergraph neural network designed for accurate drug–drug interaction (DDI) prediction using only SMILES representations. It constructs a hypergraph from molecular substructures extracted from SMILES strings, allowing the model to capture complex chemical relationships beyond pairwise similarities. HyGNN introduces an attention-based hypergraph edge encoder to generate expressive hyperedge representations and a decoder to predict interactions between drug pairs. Extensive experiments demonstrate that HyGNN achieves superior performance across multiple evaluation metrics, including F1 score, ROC-AUC, and PR-AUC, and generalizes effectively to new and unseen drugs, highlighting its robustness and practical applicability in drug discovery.

The HeTAN workflow comprises three steps: Heterogeneous Graph Construction, Encoder, and Decoder. Initially, a heterogeneous network is built with drug, target, and disease nodes connected by drug-target or drug-disease edges. Target disease connections (dashed lines) are inferred from shared drug associations. We introduce Triplet Message Passing (TMP) and Triplet-wise attention to generate node representations. Finally, using concatenated node representations, a Multi-Layer Perceptron (MLP) predicts drug-target-disease interactions.

HeTriNet: Heterogeneous Graph Triplet Attention Network

📘 DSAA'24  ·  Triplet Attention HGNN

Overview. HeTAN- a heterogeneous graph triplet attention network that models complex interactions among drugs, targets, and diseases. By introducing triplet message passing and triplet-wise attention, it captures higher-order relationships beyond pairwise interactions. This design enables HeTAN to uncover meaningful drug–target–disease associations and achieve superior performance on real-world biomedical datasets.

Topic: Advancing Sequence Data Analysis through Innovative Graph/Hypergraph Learning Models

System architecture of the proposed method. The first step is hypergraph construction, where each sequence (e.g., DNA) is a hyperedge, and the (frequent) subsequences of sequences are the nodes. The second step is the Sequence Hypergraph Attention Network, namely Seq-HyGAN, which generates the representations of sequences while giving more attention to the important subsequences and learning the label of the sequences.

Seq-HyGAN: Sequence understanding using Hypergraph Attention Network

📘 CIKM'23  ·  Sequence Analysis HyperGNN

Overview. Seq-HyGAN- sequence hypergraph attention network that models higher-order relationships among subsequences for sequence classification. By capturing both local and global dependencies through attention-based learning, it produces richer representations and achieves superior accuracy over existing approaches.

Overview of the proposed graph-based Drug Abuse detection framework. The process includes data collection and preprocessing, construction of corpus- and document-level text graphs to capture relationships, and the use of Graph Neural Networks (GNNs) for effective classification of DA-related tweets.

Drug Abuse Detection from Social Network via Graph Learning

📘 BigData'21  · GNN Text Understanding

Overview. This work detects drug abuse (DA) events from social media using graph neural networks (GNNs). By constructing text graphs from Twitter data at corpus and document levels, it captures word- and document-level relationships for effective classification. Experimental results show that GNN-based models outperform traditional machine learning and deep learning approaches in identifying DA-related content.

Topic: Dynamic Graph Neural Network for Event Understanding

An overview of Dynamic Graph Contrastive Learning, (DyGCL) architecture. We feed input graphs into the Local View Encoder and Global View Encoder. Local View encoder learns the dynamic graph representation through the dynamic node representation. Global-View Encoder learns dynamic graph representation through graph pooling and the LSTM model. After optimizing representations by contrastive learning, they are combined by an MLP layer and fed to the predictor for event prediction.

Dynamic graph contrastive learning for event prediction

📘 BigData'24  · DyGNN Event Understanding

Overview. DyGCL- Dynamic Graph Contrastive Learning framework designed for event prediction. It integrates both local and global view encoders to capture fine-grained node dynamics and the overall structural evolution of graphs. By optimizing these representations through contrastive learning and combining them via an attention mechanism, DyGCL effectively models temporal patterns and achieves superior performance on multiple real-world datasets.

Topic: Biomedical Knowledge Mining from Social Networks

Overview of the five-step framework for analyzing COVID-19’s impact on Active MAT Medicine Users (AMMUs) using Twitter data.

Effects of COVID-19 in Opioid Addiction Recovery

📘 ICMLA'21  · Social Network Text Mining

Overview. This study analyzes Twitter data to assess COVID-19’s impact on opioid treatment, revealing a 30% drop in MAT use and increased illicit opioid mentions, highlighting relapse risks and the need for continued care.