Executive Summary : | Learning from graph data is crucial for various domains, such as social networks, traffic networks, and citation networks. However, learning temporal models on dynamic graph data is challenging due to several design issues. Most existing methods are designed for static graphs, which are not suitable for continuous-time dynamic graphs. Directly extending static methods leads to suboptimal performance, necessitating customized algorithms for dynamic graphs. Discrete time dynamic graph models approximate the continuous-time dynamic graph to a series of static graph snapshots with uniform time intervals, but this compromises modeling precision. Recent methods focusing on temporal patterns in dynamic graph data often ignore topological properties and assume that events in graph data are generated under the influence of internal dynamics only. However, external events significantly influence events in a graph network, and incorporating these effects into the model can significantly improve model performance at downstream tasks. To address this problem, a continuous-time dynamic graph model is proposed, employing a neural attention-based point process framework for modeling continuous-time events and graph neural network-based models to capture the node neighborhood. The hybrid model should capture both rich temporal patterns conditioned upon complex topological patterns and incorporate external stimuli in a domain-agnostic manner. The proposed framework will incorporate external stimuli in a domain-agnostic manner, requiring no feature engineering. The contrastive learning-based approach will be explored to train the model, allowing for more accurate modeling of temporal events in dynamic graphs, impacting downstream applications like recommender systems in e-commerce and predicting upcoming communications in community networks. |