Modeling Event Propagation via Graph Biased Temporal Point Process.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Sequential Data Modeling

Background:

• Temporal point processes are crucial for modeling sequential data.
• Conventional models often overlook graph structures in event propagation.
• Existing methods struggle to capture both direct node influence and historical event patterns.

Purpose of the Study:

• To propose a novel Graph-Biased Temporal Point Process (GBTPP) model.
• To effectively integrate latent graph structures into temporal point process modeling.
• To enhance the modeling of sequential event propagation by considering graph topology.

Main Methods:

• Leveraging graph representation learning to capture node embeddings.
• Modeling direct node influence and indirect influence from event history.
• Integrating learned node embeddings as side information into event history.

Main Results:

• The proposed GBTPP model demonstrates superior performance compared to conventional methods.
• Experimental results on synthetic and real-world datasets validate the model's efficacy.
• The model successfully captures both graph structure and temporal dynamics in event propagation.

Conclusions:

• GBTPP offers a powerful approach for modeling sequential event propagation on graphs.
• Integrating graph structure significantly improves the accuracy of temporal point process models.
• The findings have implications for understanding and predicting phenomena like information diffusion.