Publications

Predicting events, ranging from political unrest to disease outbreaks and criminal activities, stands as a pivotal task in proactively addressing emerging challenges. Despite the richness of textual data as a source for event detection, it is challenging to extract contextual information from documents due to their complex structure and the dynamic evolution of events. In response to this challenge, dynamic Graph Neural Networks (GNNs) have emerged as a promising tool for capturing the intricate patterns embedded within textual data graphs. Nevertheless, many models in this domain primarily rely on local node-level representations, overlooking the essential global graph-level context. However, both node-level and graph-level representations are critical for effective event prediction. Node-level representations provide insight into the local structure, while graph-level representations offer an understanding of the global structure and the evaluation of temporal graphs. To address these challenges, in this paper, we propose a Dynamic Graph Contrastive Learning (DyGCL) method for event prediction. Our model DyGCL first employs a local view encoder to effectively capture the local dynamic structure of input graphs as the evolving node representations. Then, it performs a global view encoder to perceive the hierarchical dynamic graph representation of the input graphs. Finally, the graph representations from both encoders, optimized via contrastive learning, are combined with an attention mechanism and utilized to predict future events. Our extensive experiments demonstrate that our proposed method outperforms the state-of-the-art methods for event prediction on six real-world datasets.

Recently, Graph Neural Networks (GNNs) have emerged as a powerful technique for various graph-related tasks. Current GNN models apply different graph pooling methods that reduce the number of nodes and edges to learn the higher-order structure of the graph in a hierarchical way. However, these methods primarily rely on the one-hop neighborhood and do not consider the higher-order structure of the graph. To address this issue, in this work, we propose a multi-channel Motif-based Graph Pooling method named (MPool) that captures the higher-order graph structure with motif and also considers the local and global graph structure through a combination of selection and clustering-based pooling operations. In the first channel, we develop node selection-based graph pooling by designing a node ranking model considering the motif adjacency of nodes. In the second channel, we develop cluster-based graph pooling by designing a spectral clustering model using motif adjacency. Finally, the result of each channel is aggregated into the final graph representation. We perform extensive experiments and demonstrate that our proposed method outperforms the baseline methods for graph classification tasks on eight benchmark datasets.

We propose a novel Hypergraph Attention Network model, namely Seq-HyGAN. Our approach is built on the assumption that sequences sharing structural similarities tend to belong to the same classes, and sequences can be considered similar if they contain similar subsequences. To effectively capture the structural similarities between sequences, we represent them in a hypergraph framework, where the sequences are depicted as hyperedges that connect their respective subsequences as nodes. This construction allows us to create a single hypergraph encompassing all the sequences in the dataset. Unlike a standard graph where the degree of each edge is 2, hyperedge is degree-free; it can connect an arbitrary number of nodes. To enhance the representation of sequences and capture complex relationships among them, we introduce a novel Seq-HyGAN architecture that employs a three-level attention-based neural network.

We design a heterogeneous information network (HIN) to model relations between these entities. Afterward, we extract the rich semantic relationships among these entities using different meta-path-based topological features. An extensive set of features are fed to different classifiers.

We propose a novel Network Embedding method, NECL, to generate embedding more efficiently or effectively. Our goal is to answer the following two questions: 1) Does the network Compression significantly boost learning? 2) Does network compression improve the quality of the representation? For these goals, first, we propose a novel graph compression method based on the neighborhood similarity that compresses the input graph to a smaller graph with incorporating local proximity of its vertices into super-nodes; second, we employ the compressed graph for network embedding instead of the original large graph to bring down the embedding cost and also to capture the global structure of the original graph; third, we refine the embeddings from the compressed graph to the original graph.