Multistream-Based Marked Point Process With Decomposed Cumulative Hazard Functions.

Area of Science:

• Computational Statistics
• Machine Learning
• Data Science

Background:

• Point processes are crucial for modeling event data, requiring accurate intensity function models based on prior knowledge.
• Recent deep learning methods for temporal point processes model cumulative hazard functions (CHFs) adaptively.
• Existing deep learning approaches often overlook event mark information, limiting their applicability to diverse real-world scenarios.

Purpose of the Study:

• To propose a novel fully trainable marked point process method that incorporates event mark information.
• To develop a deep neural network architecture capable of predicting both time and mark information.
• To enhance the modeling of point processes for applications generating rich event data.

Main Methods:

• A multistream deep neural network architecture is proposed to model decomposed cumulative hazard functions (CHFs).
• The method enables simultaneous prediction of event occurrence times and their associated marks.
• Experiments were conducted using both synthetic and real-world event datasets to validate the approach.

Main Results:

• The proposed method effectively models decomposed CHFs for time and mark prediction.
• Demonstrated superior performance in capturing both temporal dynamics and event characteristics compared to existing methods.
• Validated the effectiveness of the multistream deep neural network for marked point process modeling.

Conclusions:

• The developed fully trainable marked point process method significantly advances the analysis of complex event data.
• Incorporating mark information via decomposed CHFs provides a more comprehensive understanding of event processes.
• The approach offers a powerful tool for various applications requiring detailed event data analysis.